HK1177040A

HK1177040A - Learning device, identification device, learning identification system and learning identification device

Info

Publication number: HK1177040A
Application number: HK13104021.0A
Authority: HK
Inventors: 利宪细井
Original assignee: 日本电气株式会社
Priority date: 2010-01-06
Filing date: 2010-12-24
Publication date: 2013-08-09

Description

Learning device, recognition device, learning recognition system, and learning recognition device

Technical Field

The present invention relates to a learning device that learns a learning object pattern in advance to perform object recognition, a learning method, a recording medium storing a learning program, a recognition device that inputs and recognizes the recognition object pattern, a recognition method, a recording medium storing a recognition program, and a learning recognition system that learns the learning object pattern and recognizes the recognition object pattern, a learning recognition device, a learning recognition method, and a recording medium storing a learning recognition program.

Background

Statistical pattern recognition methods are well known methods for recognizing objects from images.

The statistical pattern recognition method is a method of estimating a relationship between input and output based on some input-output pairs. More specifically, the statistical pattern recognition method is a method in which a desired input-output relationship is learned from a large amount of input data and output data, and the relationship is used for recognition.

Therefore, the statistical pattern recognition method is mainly realized by a learning process and a recognition process.

The learning process is to acquire learning parameters for the identification process by using the sample data for learning and teacher data thereof. The teacher data is a recognition result (output) indicating the correctness for the sample data (input).

Specifically, the teacher data is an arithmetic expression or a learning parameter for deriving an output value from an input value (for example, an input value and an output value corresponding to the input value), or both.

In other words, as described above, the learning process is a calculation process, the purpose of which is to acquire learning parameters for calculating output data for any input at the stage of the recognition process.

For example, in the case where the learning process uses a multilayer perceptron that is a kind of neural network, "connection weights" between nodes are acquired as learning parameters.

On the other hand, the recognition process is a process of calculating an output (recognition result) by using a learning parameter for input arbitrary data (recognition object).

In general, in order to improve the accuracy of the recognition process, a complicated feature extraction process is performed on each of a large number of learning object patterns. For example, when the feature recognition is performed, the slope, width, curvature, number of turns, and the like of the contour corresponding to the specific feature of the learning object pattern are extracted as the feature amount. Thus, in other words, the feature extraction process is a process of building another pattern from the initial pattern.

A general feature extraction process will be described below with reference to fig. 14.

Fig. 14 is a block diagram showing a general structure of performing a recognition process based on a statistical pattern recognition method.

First, as shown in fig. 14, after an arbitrary recognition object pattern is input, the preprocessing device a1 performs preliminary processing (noise removal and normalization) so that subsequent processing is more easily performed.

Next, the feature extraction means a2 extracts feature quantities (numerical values and signs) that give the particularity of the pattern from the recognition target pattern on which the preliminary processing has been performed.

For example, when the feature quantity d is extracted, the feature quantity can be represented by a feature vector expressed by the following formula.

X＝(x₁，x₂，…，x_d)

The recognition calculation means A3 inputs the feature quantity extracted by the feature extraction means a2 and determines "classification/category/class" of the recognition object.

Specifically, the recognition computing means A3 performs computation to determine whether the extracted feature amount is a specific object based on a computation method specified by learning parameters stored in advance in the dictionary storage unit a 4.

For example, the recognition calculation means a3 determines that the extracted feature amount is a specific object when the calculation result is "1", and determines not to be a specific object when the calculation result is "0". Also, the recognition computing means a3 can determine whether the extracted feature amount is a specific object based on whether the computation result is lower than a predetermined threshold.

In the following related art, in order to maintain high recognition accuracy, the feature vector x_dThe dimension of (c) needs to be equal to or greater than a preset value. Moreover, not only such a method has been used, but also other various methods have been used.

For example, in the method proposed in non-patent document 1, rectangular features are extracted from an object pattern.

In the method proposed in patent document 1, an orientation pattern indicating the distribution of orientation components on a feature pattern is established. Vertical direction components, horizontal direction components, and diagonal direction components based on the orientation pattern are extracted as an orientation feature pattern to perform feature recognition. In other words, in this method, the read-out features are reproduced by combining these directional components.

In the technique proposed in patent document 2, as for a plurality of measurement points on an image, adjacent areas having a narrow fixed shape for measurement are provided on both sides of a search line segment passing through each measurement point. With this technique, the direction of the luminance gradient vector of the image is measured at a plurality of nearby points in the area. In this technique, the concentration at each of the adjacent points can be calculated from the difference between the direction of the vector and the direction of the search line segment, the line segment concentration for the measurement point is calculated from all the concentrations, and when it is the maximum value, it is determined that the line segment information exists in the direction of the search line segment. In this technique, basic additional values that can be used at each measurement point are calculated in advance based on the concentration ratio. In this technique, when the direction of a vector is measured, one of basic additional values is selected and added for each direction of a search line segment, whereby the line segment concentration ratio of each measurement point is calculated, and when the value is maximum, line segment information is expected.

In the technique proposed in patent document 3, with respect to the feature amount of a specific object, by replacing the feature amount component corresponding to the background region with another value, a plurality of feature amount data different in background are generated from one object image, and the identification parameter is learned.

Documents of the prior art

Non-patent document

Non-patent document 1: paul Viola, Michael joints, "Rapid Object detecting a boost case of Simple Feature" Proceeding of Computer Vision and Pattern Recognition 2001, 2001

Patent document

Patent document 1: japanese patent laid-open No.2605807

Patent document 2: japanese patent application laid-open No.2005-284697

Patent document 3: japanese patent application laid-open No.2008-059110

Disclosure of Invention

Problems to be solved by the invention

However, when such a method is used, the following problems arise.

First, in the technique proposed in non-patent document 1, the feature for the learning process is a simple rectangle. Therefore, many parameters that are not useful for identification are included in the extracted feature quantity. Thus, when this technique is used, there arises a problem that a large number of feature quantities are required in order to maintain the recognition accuracy (in other words, in the learning process, higher-order calculations are necessary), and the calculation cost becomes high.

As a result, since the learning process takes a lot of time, such a problem arises in the technique that the period required until the recognition process can be performed is extended.

In the technique disclosed in patent document 1, since the features for recognition are complicated, higher-order computation is required in the recognition process, and thus, the recognition process cannot be smoothly performed.

In the techniques disclosed in patent documents 2 and 3, the processing content of the filter used for the calculation process is different from that of the present invention. Therefore, the above problems cannot be solved.

An object of the present invention is to provide a learning apparatus which avoids the use of high-dimensional feature quantities and complicated feature extraction necessary for maintaining and improving recognition accuracy when performing pattern recognition, thereby reducing the burden of calculation costs in the learning process and the recognition process and enabling smooth pattern recognition to be performed, a learning method, a recording medium storing the learning program, a recognition apparatus, a recognition method, a recording medium storing the recognition program, and a learning recognition system, a learning recognition apparatus, a learning recognition method, and a recording medium storing the learning recognition program.

Means for solving the problems

A learning apparatus according to an embodiment of the present invention includes: a gradient feature extraction unit that extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between luminance of the input learning object pattern on each coordinate and luminance of the periphery thereof; a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient directions, based on the extracted gradient feature amount, and subtracting gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; and a learning unit that acquires a learning parameter at each coordinate based on a predetermined learning algorithm using the gradient feature amount and the sum-difference feature amount.

Further, a learning method according to an embodiment of the present invention includes: extracting a gradient feature quantity including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between the luminance of the input learning object pattern on each coordinate and the luminance of the periphery thereof; calculating a predetermined sum-difference feature quantity by summing gradient intensity values conforming to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient direction, based on the extracted gradient feature quantity, and subtracting gradient intensity values conforming to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; learning parameters at each coordinate are acquired based on a predetermined learning algorithm using the gradient feature quantity and the sum-difference feature quantity.

Further, a recording medium storing a learning program according to an embodiment of the present invention causes a computer to function as: a gradient feature extraction unit that extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between luminance of the input learning object pattern on each coordinate and luminance of the periphery thereof; a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient directions, based on the extracted gradient feature amount, and subtracting gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; and a learning unit that acquires a learning parameter at each coordinate based on a predetermined learning algorithm using the gradient feature amount and the sum-difference feature amount.

Further, an identification apparatus according to an embodiment of the present invention includes: a gradient feature extraction unit that extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof at each coordinate based on a variation amount between the luminance at each coordinate of the input recognition target pattern and the luminance at the periphery thereof; a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient directions, based on the extracted gradient feature amount, and subtracting gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; and an identification unit that identifies a type to which the identification target pattern belongs from the one or more types, based on the gradient feature amount, the sum-difference feature amount, and part or all of the learning parameters acquired by a predetermined learning algorithm.

Still further, an identification method according to an embodiment of the present invention includes: extracting a gradient feature quantity including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation between the luminance of the input recognition target pattern on each coordinate and the luminance of the periphery thereof; calculating a predetermined sum-difference feature quantity by summing gradient intensity values conforming to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient direction, based on the extracted gradient feature quantity, and subtracting gradient intensity values conforming to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; the type to which the recognition object pattern belongs is identified from among the one or more types based on the gradient feature amount, the sum-difference feature amount, and part or all of the learning parameters acquired by a predetermined learning algorithm.

Further, a recording medium storing an identification program according to an embodiment of the present invention causes a computer to function as: a gradient feature extraction unit that extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof at each coordinate based on a variation amount between the luminance at each coordinate of the input recognition target pattern and the luminance at the periphery thereof; a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient directions, based on the extracted gradient feature amount, and subtracting gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; and an identification unit that identifies a type to which the identification target pattern belongs from the one or more types, based on the gradient feature amount, the sum-difference feature amount, and part or all of the learning parameters acquired by a predetermined learning algorithm.

Further, a learning identification system according to an embodiment of the present invention includes: a learning device including a gradient feature extraction unit that extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between luminance of an input learning object pattern on each coordinate and luminance of its periphery; a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient directions, based on the extracted gradient feature amount, and subtracting gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; and a learning unit that acquires a learning parameter at each coordinate based on a predetermined learning algorithm using the gradient feature amount and the sum-difference feature amount. A recognition device including a gradient feature extraction unit that extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between luminance of an input recognition target pattern on each coordinate and luminance of the periphery thereof; a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient directions, based on the extracted gradient feature amount, and subtracting gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; and an identification unit that identifies a type to which the identification target pattern belongs from among the one or more types, based on the gradient feature amount, the sum-difference feature amount, and part or all of the learning parameters acquired by the learning unit.

Still further, a learning identification apparatus including an identification unit for identifying a type to which an identification target pattern belongs from one or more types, includes: a gradient feature extraction unit that extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between luminance of the input learning object pattern on each coordinate and luminance of the periphery thereof; a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient directions, based on the extracted gradient feature amount, and subtracting gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; a learning unit that acquires a learning parameter on each coordinate based on a predetermined learning algorithm using the gradient feature quantity and the sum-difference feature quantity; the gradient feature extraction unit extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between the brightness of the input recognition target pattern on each coordinate and the brightness of the periphery thereof; the sum-difference feature extraction unit calculating a predetermined sum-difference feature amount by summing gradient intensity values for gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient direction, and subtracting gradient intensity values for gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value, based on the extracted gradient feature amount; and an identification unit that identifies a type to which the recognition target pattern belongs from among the one or more types, based on the gradient feature amount and the sum-difference feature amount calculated from the recognition target pattern, and part or all of the learning parameters acquired by the learning unit.

Further, the learning identification method according to an embodiment of the present invention includes the steps of: extracting a gradient feature quantity including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between the luminance of the input learning object pattern on each coordinate and the luminance of the periphery thereof; calculating a predetermined sum-difference feature quantity by summing gradient intensity values conforming to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient direction, based on the extracted gradient feature quantity, and subtracting gradient intensity values conforming to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; acquiring learning parameters on each coordinate based on a preset learning algorithm by using the gradient characteristic quantity and the sum-difference characteristic quantity; extracting a gradient feature quantity including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation between the luminance of the input recognition target pattern on each coordinate and the luminance of the periphery thereof; calculating a predetermined sum-difference feature quantity by summing gradient intensity values conforming to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient direction, based on the extracted gradient feature quantity, and subtracting gradient intensity values conforming to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; the type to which the recognition target pattern belongs is recognized from one or more types based on the gradient feature amount and the sum-difference feature amount calculated from the recognition target pattern, and part or all of the learning parameters acquired by the learning unit.

Further, a recording medium storing a learning identification program according to an embodiment of the present invention causes a computer to function as: a gradient feature extraction unit that extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between luminance of the input learning object pattern on each coordinate and luminance of the periphery thereof; a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient directions, based on the extracted gradient feature amount, and subtracting gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; a learning unit that acquires a learning parameter on each coordinate based on a predetermined learning algorithm using the gradient feature quantity and the sum-difference feature quantity; the gradient feature extraction unit extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between the brightness of the input recognition target pattern on each coordinate and the brightness of the periphery thereof; the sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient directions, based on the extracted gradient feature amount, and subtracting gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; and an identification unit that identifies a type to which the recognition target pattern belongs from among the one or more types, based on the gradient feature amount and the sum-difference feature amount calculated from the recognition target pattern, and part or all of the learning parameters acquired by the learning unit.

Effects of the invention

The learning device, the learning method, the recording medium storing the learning program, the recognition device, the recognition method, the recording medium storing the recognition program, and the learning recognition system, the learning recognition device, the learning recognition method, and the recording medium storing the learning recognition program can reduce the burden of calculation cost, improve recognition accuracy, and realize smooth learning process and recognition process.

Drawings

Fig. 1 is a first block diagram showing the structure of a learning apparatus according to a first embodiment of the present invention;

fig. 2 is an explanatory diagram showing gradient feature quantities used for the learning process according to the first embodiment of the invention;

fig. 3A is an explanatory diagram showing sum and difference feature quantities used for the learning process according to the first embodiment of the invention;

fig. 3B is an explanatory diagram showing sum and difference feature quantities used for the learning process according to the first embodiment of the invention;

fig. 3C is an explanatory diagram showing sum and difference feature quantities used for the learning process according to the first embodiment of the invention;

fig. 4 is a second block diagram showing the structure of a learning apparatus according to the first embodiment of the present invention;

fig. 5 is a first block diagram showing the structure of an identification apparatus according to a first embodiment of the present invention;

fig. 6 is a second block diagram showing the structure of an identification apparatus according to the first embodiment of the present invention;

fig. 7 is a flowchart showing the steps of the learning process according to the first embodiment of the present invention;

FIG. 8 is a flowchart illustrating the steps of an identification process according to a first embodiment of the present invention;

fig. 9A is a diagram showing a storage method of simply assigning gradient feature quantity data according to a second embodiment of the present invention;

fig. 9B is a diagram showing a storage method of optimally assigning gradient feature quantity data according to the second embodiment of the present invention;

fig. 9C is a diagram showing an example of a method of storing gradient feature quantities according to the second embodiment of the present invention;

fig. 10 is a first block diagram showing the structure of a learning identification system according to a third embodiment of the present invention;

fig. 11 is a second block diagram showing the structure of a learning identification system according to a third embodiment of the present invention;

fig. 12 is a first block diagram showing the structure of a learning identification apparatus according to a fourth embodiment of the present invention;

fig. 13 is a second block diagram showing the structure of a learning identification apparatus according to a fourth embodiment of the present invention;

fig. 14 is a block diagram showing the structure of a related art recognition apparatus;

Detailed Description

Embodiments of the present invention will be described below.

Here, the learning apparatus, the learning method, the identification apparatus, the identification method, the learning identification system, the learning identification apparatus, and the learning identification method in the exemplary embodiments mentioned later are realized by a method, an apparatus, or a function executed by a computer according to program instructions (for example, a learning program, an identification program, and a learning identification program). The program sends a command to each component of the computer to cause it to execute a predetermined method and to function as shown below. That is, the learning apparatus, the learning method, the identifying apparatus, the identifying method, the learning identifying system, the learning identifying apparatus, and the learning identifying method in the embodiments described later are realized by specific apparatuses in which the program and the computer work in cooperation with each other.

Further, all or part of the program is provided by, for example, a magnetic disk, an optical disk, a semiconductor memory, or any computer-readable recording medium, and the program read out from the recording medium is installed in a computer and executed. Also, the program can be directly loaded into the computer through a communication line without using a recording medium.

[ first embodiment ]

(learning device 10)

Fig. 1 is a first block diagram showing the structure of a learning apparatus according to a first embodiment of the present invention.

As shown in fig. 1, the learning apparatus 10 according to the embodiment of the present invention includes a gradient feature extraction unit 11, a sum-difference feature extraction unit 12, and a learning unit 13, and the learning apparatus 10 is connected to a storage apparatus 20.

The gradient feature extraction unit 11 extracts a gradient feature amount at each coordinate from the input learning object pattern.

The gradient feature extraction unit 11 extracts the amount of change between the luminance of the learning object pattern at each coordinate and the luminance of its periphery. The gradient feature amount is data extracted by quantizing the direction and intensity of luminance (gradient intensity value) at each coordinate based on the change amount.

Fig. 2 is an explanatory diagram intuitively showing the gradient feature amount extracted by the gradient feature extraction unit 11.

As shown in fig. 2, for example, when a photographic image of a human face is input as a learning object pattern before feature extraction, the gradient feature extraction unit 11 can divide the image into eight gradients (0 degrees, 45 degrees, 90 degrees, 135 degrees, 180 degrees, 225 degrees, 270 degrees, and 315 degrees) by using a gradient filter such as a SOBEL filter or the like and extract gradient feature amounts.

In other words, the gradient feature extraction unit 11 extracts a gradient intensity value for each of the eight gradients, which are extracted separately for each pixel. The right-hand graph of fig. 2 schematically shows the result of extracting gradient intensity values of all pixels of the object image (left-hand graph of fig. 2).

The gradient feature extraction unit 11 temporarily stores the extracted gradient feature in the feature temporary storage unit 22 of the storage device 20.

As shown in fig. 1, the sum-difference feature extraction unit 12 obtains gradient feature quantities from the feature quantity temporary storage unit 22 and calculates sum-difference feature quantities. The sum-difference feature extraction unit 12 can receive the gradient feature amount from the gradient feature extraction unit 11.

Specifically, the sum and difference feature extraction unit 12 adds gradient intensity values that conform to gradient directions included in the "predetermined gradient range" to specific coordinates of the learning object pattern. The sum-difference feature extraction unit 12 calculates a sum-difference feature quantity by subtracting gradient intensity values that conform to gradient directions included in "other gradient ranges" adjacent to the gradient range.

Here, the "predetermined gradient range" is a range including gradient feature quantities (gradient intensity values) conforming to two or more gradient directions when the entire range including all the effective gradients, that is, 360 degrees, is quantized into four or more gradient directions.

The "other gradient range" is a range adjacent to the "predetermined gradient range", and includes gradient feature quantities conforming to gradient directions whose number is the same as the number of gradient directions included in the "predetermined gradient range".

Fig. 3A is a diagram schematically showing an example of visually indicating gradient intensity values for each gradient direction.

Fig. 3B is an explanatory diagram showing a method of explaining a feature range as a calculation target in fig. 3C. The range represented by the white rectangle indicates the range of the feature value as the addition object. The range represented by the black rectangle indicates the range of the feature value as the subtraction object.

Fig. 3C is an explanatory diagram illustrating the sum-difference feature quantity when the number of gradient directions to be quantized is 8.

In fig. 3C, the range (width) in the gradient direction to be added or subtracted is indicated at the head of a row, and the arrangement of the range in the gradient direction to be added or subtracted is indicated at the head of a column.

For example, when the width is "2" and θ is 45 degrees, there are 8 permutations of addition (+) and subtraction (-) of gradient intensity values: (1)"(+): 4 θ and 3 θ "(-) -c: 2 θ and θ "; (2)"(+): 5 θ and 4 θ "," (-) -: 3 θ and 2 θ "; (3)"(+): 6 θ and 5 θ "," (-) -: 4 θ and 3 θ "; (4)"(+): 7 θ and 6 θ "," (-) -: 5 θ and 4 θ "; (5)"(+): 0 and 7 θ "," (-) -: 6 θ and 5 θ "; (6)"(+): θ and 0 "," (-) -: 7 θ and 6 θ "; (7)"(+): 2 θ and θ "," (-) -: 0 and 7 θ "; and (8) "(+): 3 θ and 2 θ "and" (-) -: theta and 0'.

The sum-difference feature extraction unit 12 calculates a sum-difference feature quantity by performing addition and subtraction on the gradient intensity values according to each of these combinations.

For example, when using E (θ)_n) Indicating compliance with an arbitrary gradient direction (θ)_n) The sum and difference feature quantities of the array (1) can be calculated by the following method. That is, the sum-difference feature extraction unit 12 can acquire one of these sum-difference feature amounts for the pixel by performing the calculation process E (4 θ) + E (3 θ) - (E (2 θ) + E (θ)). The sum-difference feature extraction unit 12 can calculate a required sum-difference feature quantity with respect to the pixels by performing the calculation process on the arrangements (1) to (8). In other words, in this case, the sum and difference feature extraction unit 12 calculates 8 sum and difference feature amounts with respect to one pixel.

In addition, the sum and difference feature extraction unit 12 performs the method for all pixels of the learning object pattern.

Further, the "gradient range" described above may be set or changed according to an input operation by the user.

Specifically, the learning device 10 can arbitrarily set the "predetermined gradient range" or the "other gradient range" (the function of the gradient feature number setting device of the present invention) by setting the number of gradient directions included in the gradient range.

For example, when the width is set to "3", the sum and difference feature extraction unit 12 calculates 16 sum and difference feature amounts: 8 sum and difference feature quantities obtained by performing addition and/or subtraction of the gradient intensity values with the width of "3", and 8 sum and difference feature quantities calculated by performing addition/subtraction of the gradient intensity values with the width of "2".

Similarly, when the width is set to "4", the sum and difference feature extraction unit 12 acquires 24 sum and difference feature amounts: 8 sum and difference feature quantities obtained by performing addition and/or subtraction of the gradient intensity values with the width of "4", 8 sum and difference feature quantities calculated by performing addition and/or subtraction of the gradient intensity values with the width of "3", and 8 sum and difference feature quantities calculated by performing addition/subtraction of the gradient intensity values with the width of "2".

Thus, the sum and difference feature extraction unit 12 can adjust the number of sum and difference feature amounts used for the learning process by changing the range (width) for addition and/or subtraction of the calculated sum and difference feature amounts.

Further, the sum-difference feature extraction unit 12 calculates sum-difference feature quantities for all pixels, and temporarily stores the calculated sum-difference feature quantities in the feature quantity temporary storage unit 22 of the storage device 20.

As shown in fig. 1, the learning unit 13 acquires the gradient feature amount and the sum-difference feature amount from the feature amount temporary storage unit 22, and obtains the learning parameters of the recognition process based on a predetermined learning algorithm. The learning unit 13 receives these gradient feature amounts and sum-difference feature amounts through the gradient feature extraction unit 11 and the sum-difference feature extraction unit 12.

Further, as shown in fig. 1, the learning unit 13 can acquire the learning parameters by using the gradient feature amount and the sum-difference feature amount and predetermined teacher data.

The teacher data is data indicating a correct recognition result (output) for the sample data (input). Specifically, the teacher data is an arithmetic expression or a learning parameter for deriving an output value from an input value, or both.

Here, the learning unit 13 can acquire not only the learning parameters by using a specific learning algorithm but also various learning algorithms.

For example, in the case of a multilayer perceptron using a kind of neural network, the learning unit 13 acquires "connection weights" between the nodes as learning parameters.

When Generalized Learning Vector Quantization (GLVQ) is used as the learning algorithm, the learning unit 13 may acquire a "reference vector (prototype)" as the learning parameter.

When a Support Vector Machine (SVM) is used as the learning algorithm, the learning unit 13 can acquire a selected "support vector" as a learning parameter.

The learning unit 13 stores the acquired learning parameters in the dictionary storage unit 21.

The learning parameters stored in the dictionary storage unit 21 are used for a recognition process described later.

Further, as shown in fig. 4, the learning device 10 may have a structure in which the storage device 20 is not included.

In this case, the gradient feature extraction unit 11 can directly output the extracted gradient feature amount to the sum and difference feature extraction unit 12 and the learning unit 13. The sum-difference feature extraction unit 12 can directly output the calculated sum-difference feature amount to the learning unit 13, and send the gradient feature amount received from the gradient feature extraction unit 11 to the learning unit 13.

The learning unit 13 outputs learning parameters acquired based on the gradient feature amount and the sum-difference feature amount (including teacher data when teacher data is used) to the identification means.

(identification means 30)

As shown in fig. 5, the recognition device 30 includes a gradient feature extraction unit 31, a difference feature amount extraction unit 32, and a recognition unit 33.

The gradient feature extraction unit 31 extracts a gradient feature amount at each coordinate from the input recognition target pattern.

The gradient feature extraction unit 31 extracts the amount of change between the luminance of the recognition target pattern at each coordinate and the luminance of the periphery thereof. The gradient feature amount is data extracted by quantizing the direction and intensity of luminance (gradient intensity value) at each coordinate based on the change amount.

That is, the gradient feature extraction unit 31 performs the same operation as the gradient feature extraction unit 11 of the learning device 10.

Further, the gradient feature extraction unit 31 can extract only a specific gradient feature amount determined to be useful for the recognition process. For example, the gradient feature extraction unit 31 excludes the gradient feature amount determined to be useless for the recognition process by a statistical method or the like, and attaches an identification code useless for the recognition process to predetermined coordinates or pixels of the recognition target pattern. Thus, it is possible to eliminate extraction of gradient feature quantities that are not useful for recognition.

The gradient feature extraction unit 31 temporarily stores the extracted gradient feature amounts in the feature amount temporary storage unit 42 of the storage device 40.

As shown in fig. 5, the sum-difference feature extraction unit 32 acquires gradient feature amounts from the feature amount temporary storage unit 42, and calculates predetermined sum-difference feature amounts. The sum-difference feature extraction unit 32 can receive the gradient feature amount directly from the gradient feature extraction unit 31.

That is, the sum and difference feature extraction unit 32 performs the same operation as the sum and difference feature extraction unit 12 of the learning device 10.

For example, when the "predetermined gradient range" and the "other gradient range" in the sum-difference feature extraction unit 12 of the learning device 10 are set to "3", in general, the sum-difference feature extraction unit 32 of the recognition device 30 calculates the sum-difference feature quantity by using the gradient feature quantities included in the gradient ranges having the same width. Specifically, in this case, the sum-difference feature extraction unit 32 sums up gradient intensity values that conform to 3 gradient directions and subtracts gradient intensity values that conform to 3 other gradient directions from the calculated sum value to acquire a sum-difference feature quantity. The sum-difference feature extraction unit 32 sums up gradient intensities that correspond to two gradient directions and subtracts gradient intensity values that correspond to two other gradient directions from the calculated sum value to obtain a sum-difference feature quantity.

However, as a result of the learning process of the learning device 10, when the gradient feature amount is determined to be useless for identification when the width of the gradient range is "3", the sum-difference feature extraction unit 32 can acquire the sum-difference feature amount based only on the gradient range having the width of "2". In other words, in this case, the sum-difference feature extraction unit 32 sums gradient intensity values that conform to two gradient directions and subtracts gradient intensity values that conform to two other gradient directions from the calculated sum value to acquire a sum-difference feature quantity.

Further, by using the same method as the method performed by the gradient feature extraction unit 31 described earlier, the sum-difference feature extraction unit 32 can acquire only the specific sum-difference feature quantities determined to be useful for identification.

As a result, the learning device 10 according to the exemplary embodiment can reduce the calculation cost of recognition and realize a quick recognition process.

Further, the sum-difference feature extraction unit 32 temporarily stores the calculated sum-difference feature quantity in the feature quantity temporary storage unit 42 of the storage device 40.

As shown in fig. 5, the recognition unit 33 obtains the gradient feature amount and the sum-difference feature amount from the feature amount temporary storage unit 42. The recognition unit 33 obtains the learning parameters from the dictionary storage unit 41, and based on these learning parameters, outputs information indicating a type to which the recognition target pattern belongs among one or more types (predetermined classification/category/class or the like) as a recognition result.

For example, the identifying unit 33 acquires euclidean distances between a feature vector composed of a gradient feature amount and a sum-difference feature amount extracted from the identification target pattern and reference vectors of a plurality of learning parameters, identifies a class/category/class to which the identification target pattern belongs, and to which the reference vector whose euclidean distance is the shortest distance is assigned.

Here, the recognition unit 33 can extract only the learning parameters determined to be useful for recognition from the learning parameters stored in the dictionary storage unit 41 and use them in the recognition process. For example, the recognition unit 33 can implement the method by attaching an identification code to a learning parameter determined to be useful for recognition in advance based on a recognition result extracted by an actual recognition process.

Further, as shown in fig. 6, the identification device 30 has a structure in which the storage device 40 may not be included.

In this case, the gradient feature extraction unit 31 can directly output the extracted gradient feature amount to the sum-difference feature extraction unit 32 or the identification unit 33. The sum-difference feature extraction unit 32 can directly output the calculated sum-difference feature quantity to the recognition unit 33. The sum-difference feature extraction unit 32 can transmit the gradient feature amount received from the gradient feature extraction unit 31 to the learning unit 33.

The learning unit 13 outputs a recognition result based on the gradient feature amount, the sum difference feature amount, and the learning parameters and the like received from a learning device (not shown).

(storage means 20(40))

The storage device 20 shown in fig. 1 and the storage device 40 shown in fig. 5 are storage devices composed of a hard disk, a semiconductor memory, or the like.

As shown in these figures, the storage device 20(30) includes a dictionary storage unit 21(41) and a feature quantity temporary storage unit 22 (42).

The dictionary storage unit 21(41) stores the learning parameters acquired by the learning process. The learning parameters stored in the dictionary storage unit 21(41) are used for the recognition process.

The feature quantity temporary storage unit 22 stores the gradient feature quantity extracted by the gradient feature extraction unit 11 and the sum-difference feature quantity calculated by the sum-difference feature extraction unit 12 in the learning device 10.

The feature quantity temporary storage unit 42 stores the gradient feature quantity extracted by the gradient feature extraction unit 31 and the sum-difference feature quantity calculated by the sum-difference feature extraction unit 32 in the identification device 30.

Next, a learning method of the learning device having the above-described structure and a recognition method of the recognition device having the above-described structure are described with reference to fig. 7 and 8.

Fig. 7 is a flowchart illustrating a learning method according to an embodiment.

Fig. 8 is a flowchart illustrating an identification method according to an embodiment.

(learning method)

As shown in fig. 7, in the learning method according to the exemplary embodiment, first, the learning apparatus 10 inputs a learning object pattern through an input unit (not shown) (S101).

Next, the gradient feature extraction unit 11 extracts a gradient feature amount from the input learning object pattern (S102).

Specifically, the gradient feature extraction unit 11 extracts a gradient feature amount composed of the gradient direction on each coordinate and its gradient intensity value based on the amount of change between the luminance of the input learning object pattern on each coordinate and the luminance of its periphery.

The gradient feature extraction unit 11 stores the extracted gradient feature in the feature amount temporary storage unit 22 of the storage device 20.

Next, the sum-difference feature extraction unit 12 obtains the gradient feature amount from the feature amount temporary storage unit 22, and calculates the sum-difference feature amount (S103).

Specifically, based on the gradient feature amount extracted from the learning object pattern, the sum difference feature extraction unit 12 sums gradient intensity values that conform to gradient directions included in a predetermined gradient range. The sum-difference feature extraction unit 12 calculates a sum-difference feature quantity by subtracting gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the gradient range.

The sum-difference feature extraction unit 12 performs the above-described method on all pixels in the learning object pattern.

Further, the sum and difference feature extraction unit 12 stores the acquired sum and difference feature quantities in the feature quantity temporary storage unit 22.

Next, the learning device 10 determines whether there are other learning object patterns to be learned (S104).

When there are other learning object patterns to be learned (S104: yes), the learning device 10 performs a similar method on the other learning object patterns (S101 to S103). That is, the learning device 10 repeats the methods of S101 to S103 the same number of times as the number of learning target patterns.

When there are no other learning object patterns to be learned (S104: no), the learning unit 13 performs a learning process (S105).

Specifically, the learning unit 13 acquires the learning parameters based on a predetermined learning algorithm using the gradient feature amount and the sum-difference feature amount taken from the feature amount temporary storage unit 22.

Further, at that time, the learning unit 13 can perform the learning process by adding predetermined teacher data.

In the present exemplary embodiment, the learning unit 13 stores the acquired learning parameters in the dictionary storage unit 41 of the storage device 40 connected to the recognition device 30.

Here, the steps of the above-described learning process can be arbitrarily changed.

For example, when a learning process is performed in which learning object patterns are sequentially input as in Generalized Learning Vector Quantization (GLVQ), the processing order of steps S104 and S105 can be interchanged. That is, the following series of steps may be repeatedly performed according to the number of learning patterns: learning object patterns are input- > gradient feature quantity extraction- > extraction and difference feature quantity- > a learning process is performed.

(identification method)

As shown in fig. 8, in the recognition method according to the exemplary embodiment, first, the recognition apparatus 30 inputs a recognition object pattern through an input unit (not shown) (S201).

Next, the gradient feature extraction unit 31 extracts a gradient feature amount from the input recognition target pattern (S202).

Specifically, the gradient feature extraction unit 31 extracts a gradient feature amount composed of the gradient direction at each coordinate and its gradient intensity value based on the amount of change between the luminance of the input recognition target pattern at each coordinate and the luminance of its periphery.

The gradient feature extraction unit 31 stores the extracted gradient feature in the feature amount temporary storage unit 42 of the storage device 40.

Next, the sum-difference feature extraction unit 32 obtains the gradient feature amount from the feature amount temporary storage unit 42, and calculates the sum-difference feature amount (S203).

Specifically, based on the gradient feature amount extracted from the recognition target pattern, the sum difference feature extraction unit 32 sums gradient intensity values that conform to gradient directions included in a predetermined gradient range. The sum-difference feature extraction unit 32 calculates a sum-difference feature quantity by subtracting gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the gradient range.

The sum-difference feature extraction unit 32 performs the above-described method on all pixels in the recognition target pattern.

Further, the sum and difference feature extraction unit 32 stores the calculated sum and difference feature quantities in the feature quantity temporary storage unit 42.

Next, the recognition unit 33 acquires the gradient feature amount and the sum-difference feature amount from the feature amount temporary storage unit 42, and acquires the learning parameters stored in the step of the learning process from the dictionary storage unit 41. Based on these data, the recognition unit 33 recognizes the type to which the recognition object pattern belongs from the one or more types (S204).

The recognition result may be output by a display, printing, or the like.

As described above, the learning device 10 of the exemplary embodiment extracts the gradient feature amount at each coordinate based on the input learning object pattern, extracts the sum and difference feature amounts based on the gradient feature amount, and acquires the learning parameter based on the predetermined learning algorithm.

For this reason, the learning device 10 of the exemplary embodiment can extract the learning parameter based on the specific feature amount by performing a simple calculation process such as addition and subtraction.

Therefore, the learning device 10 of the exemplary embodiment can reduce the calculation cost of the learning process and realize a smooth learning process.

Further, the recognition device 30 of this exemplary embodiment acquires the gradient feature amount at each coordinate based on the input recognition target pattern, acquires the sum and difference feature amounts based on the gradient feature amounts, compares these feature amounts with the learning parameters extracted in advance (by the aforementioned learning device or the like), and thereby performs recognition of the recognition target pattern.

Therefore, the recognition device 30 of the exemplary embodiment can reduce the calculation cost of the recognition process and smoothly obtain the recognition result.

Thus, by using the learning device 10 or the recognition device 30 according to the exemplary embodiment, although the learning parameter is a specific feature amount, by performing a simple calculation process such as addition or subtraction, it is possible to efficiently obtain the learning parameter useful for recognition and smoothly calculate the recognition result.

That is, the learning device 10 or the recognition device 30 according to this exemplary embodiment does not need to extract a large number of feature amounts and learning parameters in the learning process to ensure the accuracy of recognition. In the learning device 10 or the recognition device 30 according to the exemplary embodiment, since a complicated calculation process is not required to perform the recognition process, the overall calculation cost is reduced.

In addition, in the learning apparatus 10 or the recognition apparatus 30 according to the exemplary embodiment, the amounts of gradient features required for the learning process and the recognition process can be adjusted.

For example, when a wide gradient range is set in the learning device 10 or the recognition device 30 according to the exemplary embodiment, many learning parameters can be acquired when the learning process is performed, so that the recognition accuracy in the recognition process can be further improved.

On the other hand, when a narrow gradient range is set in the learning device 10 or the recognition device 30 according to the embodiment, the calculation cost of performing the learning process and the recognition process can be reduced.

Thus, since the user can freely change the gradient range, the user can adjust the balance between the recognition accuracy and the calculation cost.

Further, the recognition device 30 of this exemplary embodiment can perform a recognition process in which only the learning parameters required for the recognition process are used among the learning parameters acquired in the learning process.

Specifically, in the identifying device 30 of this exemplary embodiment, the gradient feature extracting unit 31 extracts gradient feature amounts for identification, the sum difference feature extracting unit 32 calculates only sum difference feature amounts for identification, and the identifying unit 33 compares these feature amounts with learning parameters, thereby performing the identifying process.

Further, the recognition device 30 of the embodiment can take out only the learning parameters for recognition from the learning parameters stored in the dictionary storage unit and use them for comparison with the gradient feature amount and the sum-difference feature amount.

As a result, the recognition device 30 of the embodiment can reduce the calculation cost of the recognition process and realize a smooth recognition process.

[ second embodiment ]

Next, a learning device according to a second embodiment of the present invention will be described.

The learning device 10 according to this exemplary embodiment is characterized by a storage method used when the gradient feature amount is stored in the temporary storage unit 22. Other structures, operations/effects, and the like of the second embodiment are the same as those of the first embodiment. Therefore, detailed description will be omitted.

A method of storing gradient feature quantities according to this exemplary embodiment will be described hereinafter with reference to fig. 9A to 9C.

Fig. 9A to 9C are explanatory diagrams showing a method of storing gradient feature quantities according to this exemplary embodiment.

When the feature quantity escrow unit 22 in the learning apparatus according to the present exemplary embodiment performs the learning process, it stores, in the same storage area or in a neighboring area, respectively, the gradient feature quantity used to calculate an addition or subtraction object of the sum and difference feature quantities, and the gradient intensity value thereof extracted by the gradient feature extraction unit 11, among the gradient feature quantities composed of the gradient directions of the luminance at each coordinate.

That is, the feature amount temporary storage unit 22 does not store gradient feature amounts extracted for specific pixels of the learning object pattern in a hierarchical manner or an unordered manner (refer to fig. 9A), but stores these gradient feature amounts such that the calculation objects of these sum and difference feature amounts are arranged at the same or adjacent addresses (regions) (refer to fig. 9B).

For example, in the example of 9C, the feature amount temporary storage unit 22 stores the gradient intensity values of the respective gradient directions (directions D1 to Dn) of the pixel 001 in the storage area of the #001 address.

Further, the feature quantity temporary storage unit 22 stores the gradient intensity values of the respective gradient directions of the pixel 002 (the same operation is performed to the pixel N (N is the number of corresponding pixels of the object pattern)) in the storage area of the #002 address.

That is, the feature amount temporary storage unit 22 stores the sum and difference feature amounts so that the calculation targets of the sum and difference feature amounts are stored in the storage areas whose addresses are adjacent to (or the same as) each other in the storage areas of the memory.

Therefore, the learning device 10 of this exemplary embodiment can perform addition and subtraction of gradient intensity values at high speed when performing the sum and difference feature extraction process, as compared with the case of storing gradient feature amounts out of order, so that the period (time) required to perform the learning process can be shortened.

Further, in the exemplary embodiment, the description has been given with respect to the learning apparatus 10. However, since the structures of both the recognition device 30 and the learning device 10 are similar to each other, the operation/effect of the recognition device 30 is similar to that of the learning device 10.

[ third embodiment ]

A learning identification system according to a third embodiment of the present invention will be described below with reference to fig. 10 and 11.

Fig. 10 and 11 are block diagrams showing the structure of the learning recognition system according to this exemplary embodiment.

As shown in these figures, the learning identification system 90 according to this exemplary embodiment includes a learning device 10 and an identification device 30.

Therefore, the operation/effect of the learning identification system 90 according to this exemplary embodiment is similar to the operation/effect of the learning device 10 and the identification device 30 shown in the first embodiment. In addition, since the learning identification system 90 is composed of the learning device 10 and the identification device 30, it can be provided integrally, thereby having good applicability.

Specifically, when the learning process and the recognition process are performed in real time, the learning recognition system 90a shown in fig. 10 may be used. For example, in an embodiment in which the learning device 10 continuously inputs a learning pattern and teacher data and performs a learning process and the recognition device 30 simultaneously performs a recognition process of a recognition target pattern, the learning recognition system 90a may be applied.

In the learning identification system 90b shown in fig. 11, the learning device 10 and the identification device 30 share the storage device 20. Therefore, since the learning device 10 and the recognition device 30 can be configured not to include a storage device therein, the cost of the learning recognition system 90b can be reduced as compared with the case where the learning device 10 and the recognition device 30 include a storage device, respectively.

[ fourth embodiment ]

A learning identification device according to a fourth embodiment of the present invention will be described below with reference to fig. 12 and 13.

Fig. 12 and 13 are block diagrams showing the structure of the learning identification device according to this exemplary embodiment.

The learning identification device 100 according to this exemplary embodiment integrates the learning device 10 and the identification device 20.

Specifically, the learning identification device 100a shown in fig. 12 corresponds to the learning identification system 90a shown in fig. 10.

The learning identification device 100b shown in fig. 13 corresponds to the learning identification system 90b shown in fig. 11.

Therefore, other structures, operations/effects, and the like of the fourth embodiment are the same as those of the third embodiment.

However, in the learning identification apparatus 100 according to this exemplary embodiment, the gradient feature extraction unit and the sum-difference feature extraction unit are shared by the learning process and the identification process. This is different from the foregoing embodiment.

Therefore, the learning identification device 100 according to this exemplary embodiment can further reduce the cost as compared with the learning identification system according to the third embodiment.

By the illustrated embodiments, the description is given about the learning apparatus, the learning method, the recording medium storing the learning program, the recognition apparatus, the recognition method, and the recording medium storing the recognition program of the present invention, and the learning recognition system, the learning recognition apparatus, the learning recognition method, and the recording medium storing the learning recognition program. However, the apparatus, system, method, and recording medium storing a program according to the present invention are not limited to the foregoing embodiments, and needless to say, various changes can be made within the scope of the present invention.

For example, not only the image itself but also, for example, data after predetermined preprocessing (refer to a1 of fig. 14) is performed or data compressed by predetermined reduction processing or the like can be used as the learning object pattern and the recognition object pattern input in each embodiment.

Data on which image processing such as columnar quantization is performed on the object image can be used as the learning object pattern and the recognition object pattern. In this case, the learning apparatus, the learning method, the recording medium storing the learning program, the recognition apparatus, the recognition method, and the recording medium storing the recognition program, and the learning recognition system, the learning recognition apparatus, the learning recognition method, and the recording medium storing the learning recognition program of the present invention can suppress the influence of the brightness generated at the time of photographing. Therefore, the recognition performance can be further improved.

The invention reduces the burden of calculation cost, and has the effects of improving the recognition accuracy and smoothly executing the learning process and the recognition process. This is one example of the effect of the present invention.

The present application requests priority from japanese patent application No.2010-001265, filed on 6/1/2010, the contents of which are incorporated herein by reference in their entirety.

Industrial applicability

The present invention is applicable to object recognition such as feature recognition in images, face recognition, personal authentication using a human face or body, person detection in images of monitoring cameras, defect detection of factory production lines, and the like.

Description of the symbols

10 learning device

11 gradient feature extraction unit

12 sum and difference feature extraction unit

13 learning unit

20, 40 storage device

21, 41 dictionary memory unit

22, 42 characteristic quantity temporary storage unit

30 identification device

31 gradient feature extraction unit

32 sum and difference feature extraction unit

33 identification unit

90 learning identification system

100 learning identification device

Claims

1. A learning apparatus characterized by comprising:

a gradient feature extraction unit that extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between luminance of the input learning object pattern on each coordinate and luminance of the periphery thereof;

a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to gradient directions included in a predetermined gradient range indicating a range of predetermined gradient directions, based on the extracted gradient feature amount, and subtracting the gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum;

a learning unit that acquires a learning parameter on each coordinate based on a predetermined learning algorithm using the gradient feature amount and the sum-difference feature amount.

2. The learning apparatus according to claim 1, characterized in that:

the predetermined gradient range is a range including gradient feature quantities conforming to two or more quantized gradient directions among four or more gradient directions into which the entire range including all the available gradient directions is quantized;

the other gradient range is a range adjacent to the gradient range, which includes the gradient feature quantity conforming to the quantized gradient directions whose number is the same as the number of the quantized gradient directions included in the gradient range.

3. The learning apparatus according to claim 2, characterized in that:

the sum-difference feature quantity extraction unit sums, of gradient intensity values conforming to gradient directions included in a predetermined gradient range, gradient intensity values of a number of units smaller than the number of quantized gradient directions included in the gradient range, subtracts the gradient intensity values of the number of units included in the other gradient range from the calculated sum value.

4. A learning apparatus according to claim 2 or 3, characterized in that:

gradient feature quantity setting means is included for setting the number of the quantized gradient directions included in the preset gradient range and/or the other gradient range in accordance with an input operation.

5. The learning apparatus according to any one of claims 1 to 4, characterized in that:

the learning unit acquires the learning parameters using predetermined teacher data corresponding to the gradient feature amounts, the sum-difference feature amounts, and the predetermined learning algorithm, based on these feature amounts.

6. The learning apparatus according to any one of claims 1 to 5, characterized in that:

the learning apparatus includes a feature quantity temporary storage unit for storing a gradient feature quantity including a gradient direction and a gradient intensity value thereof at each coordinate in the learning object pattern, and

the sum-difference feature extraction unit calculates a sum-difference feature amount by summing gradient intensity values conforming to gradient directions included in a predetermined gradient range and subtracting the gradient intensity values conforming to the gradient directions included in the other gradient ranges adjacent to the predetermined gradient range, based on the gradient feature amount taken from the feature amount temporary storage unit.

7. The learning apparatus according to claim 6, characterized in that:

the feature quantity temporary storage unit stores, in the same or adjacent area, gradient feature quantities conforming to one coordinate among gradient feature quantities including a gradient direction and a gradient intensity value thereof on each coordinate of the learning object pattern.

8. A learning method characterized by comprising the steps of:

extracting a gradient feature quantity including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation between the luminance of the input learning object pattern on each coordinate and the luminance of the periphery thereof;

calculating a predetermined sum-difference feature quantity by summing the gradient intensity values conforming to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient direction, based on the extracted gradient feature quantity, and subtracting the gradient intensity values conforming to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value;

learning parameters at each coordinate are acquired based on a predetermined learning algorithm using the gradient feature quantity and the sum-difference feature quantity.

9. The learning method according to claim 8, characterized in that:

the predetermined gradient range is a range in which gradient feature quantities conforming to two or more quantized gradient directions are included in four or more gradient directions into which the entire range including all the available gradient directions is quantized;

10. A recording medium storing a learning program, characterized by causing a computer to function as:

a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing the gradient intensity values conforming to the gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient direction, and subtracting the gradient intensity values conforming to the gradient directions included in the other gradient ranges adjacent to the predetermined gradient range from the calculated sum value, based on the extracted gradient feature amount; and

and a learning unit that acquires a learning parameter on each coordinate based on a predetermined learning algorithm using the gradient feature amount and the sum-difference feature amount.

11. The recording medium storing the learning program according to claim 10, wherein:

the other gradient range is a range adjacent to the gradient range, which includes gradient feature quantities conforming to the quantized gradient directions whose number is the same as the number of the gradient directions included in the gradient range.

12. An identification device, comprising:

a gradient feature extraction unit that extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between luminance of the input recognition target pattern on each coordinate and luminance of the periphery thereof;

a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing the gradient intensity values conforming to the gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient directions, and subtracting the gradient intensity values conforming to the gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value, based on the extracted gradient feature amount; and

and an identification unit that identifies a type to which the identification target pattern belongs from among one or more types, based on the gradient feature amount, the sum-difference feature amount, and part or all of the learning parameters acquired by a predetermined learning algorithm.

13. The identification device of claim 12, wherein:

the predetermined gradient range is a range in which the gradient feature quantity conforming to two or more quantized gradient directions is included in four or more gradient directions into which the entire range including all the available gradient directions is quantized;

the other gradient range is a range adjacent to the gradient range, which includes the gradient feature quantity conforming to the quantized gradient directions whose number is the same as the number of the gradient directions included in the gradient range.

14. The identification device of claim 13, wherein:

15. The identification device according to claim 13 or 14, wherein:

16. An identification device as claimed in any one of claims 12 to 15, characterized in that:

the recognition apparatus includes a feature quantity temporary storage unit for storing the gradient feature quantity including the gradient direction and the gradient intensity value thereof at each coordinate of the recognition object pattern, and

the sum-difference feature extraction unit calculates a sum-difference feature amount by summing the gradient intensity values conforming to the gradient directions included in a predetermined gradient range and subtracting the gradient intensity values conforming to the gradient directions included in the other gradient ranges adjacent to the predetermined gradient range, based on the gradient feature amount taken from the feature amount temporary storage unit.

17. The identification device of claim 16, wherein:

the feature quantity temporary storage unit stores, in the same or an adjacent area, gradient feature quantities conforming to one coordinate among the gradient feature quantities including the gradient direction and the gradient intensity value thereof on each coordinate of the identification target pattern.

18. An identification device as claimed in any one of claims 12 to 17, characterized in that:

the recognition device includes a dictionary storage unit that stores, for a predetermined learning object parameter, a learning parameter obtained based on a predetermined learning algorithm using the gradient feature quantity and the sum-difference feature quantity;

the recognition unit recognizes a type to which a recognition target pattern belongs from one or more types based on the gradient feature amount, the sum-difference feature amount, and part or all of the learning parameters stored in the dictionary storage unit.

19. The identification device of claim 18, wherein:

the recognition unit recognizes a type to which the recognition target pattern belongs from one or more types based on the gradient feature amount, the sum-difference feature amount, and a learning parameter determined to be useful for a recognition process from among the learning parameters stored in the dictionary storage unit.

20. An identification device as claimed in any one of claims 12 to 19, wherein:

the gradient feature extraction unit extracts a specific gradient feature amount determined by the identification unit to be useful for the identification process.

21. An identification device as claimed in any one of claims 12 to 20, wherein:

the sum-difference feature extraction unit calculates a specific sum-difference feature amount determined by the recognition unit to be useful for the recognition process.

22. An identification method, characterized by comprising the steps of:

extracting a gradient feature quantity including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation between the luminance of the input recognition target pattern on each coordinate and the luminance of the periphery thereof;

calculating a predetermined sum-difference feature quantity by summing gradient intensity values conforming to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient direction, based on the extracted gradient feature quantity, and subtracting the gradient intensity values conforming to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value; and

the type to which the recognition target pattern belongs is identified from one or more types based on the gradient feature amount, the sum-difference feature amount, and part or all of the learning parameters acquired by a predetermined learning algorithm.

23. The identification method of claim 22, wherein:

24. A recording medium storing an identification program, characterized by causing a computer to function as:

a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing the gradient intensity values conforming to the gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient directions, and subtracting the gradient intensity values conforming to the gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value, based on the extracted gradient feature amount;

an identifying unit that identifies a type to which the identification target pattern belongs from one or more types based on the gradient feature amount, the sum-difference feature amount, and part or all of the learning parameters acquired by the predetermined learning algorithm.

25. The recording medium storing the identification program as set forth in claim 24, wherein:

the predetermined gradient range is a range including the gradient feature quantity conforming to two or more quantized gradient directions among four or more gradient directions into which the entire range including all the available gradient directions is quantized;

26. A learning identification system, comprising:

a learning device, comprising:

a learning unit that acquires a learning parameter on each coordinate based on a predetermined learning algorithm using the gradient feature amount and the sum-difference feature amount; and

an identification device, comprising:

a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values conforming to the gradient directions included in a predetermined gradient range based on the extracted gradient feature amount, and subtracting the gradient intensity values conforming to the gradient directions included in the other gradient range adjacent to the predetermined gradient range from the calculated sum value; and

an identifying unit that identifies a type to which the recognition target pattern belongs from one or more types based on the gradient feature amount, the sum-difference feature amount, and part or all of the learning parameters acquired by the learning unit.

27. The learning identification system of claim 26, wherein:

28. A learning identification apparatus including an identification unit for identifying a type to which an identification target pattern belongs from one or more types, characterized in that:

the learning identification device includes:

a sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to gradient directions included in a predetermined gradient range indicating a range of the predetermined gradient direction, based on the extracted gradient feature amount, and subtracting the gradient intensity values that conform to gradient directions included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value;

a learning unit that acquires a learning parameter on each coordinate based on a predetermined learning algorithm using the gradient feature amount and the sum-difference feature amount;

the gradient feature extraction means extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof at each coordinate based on a variation amount between the luminance at each coordinate of the input recognition target pattern and the luminance at the periphery thereof;

the sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing gradient intensity values that conform to the gradient directions included in a predetermined gradient range, based on the extracted gradient feature amount, and subtracting the gradient intensity values that conform to the gradient directions included in the other gradient range adjacent to the predetermined gradient range from the calculated sum value;

and an identification unit that identifies a type to which the recognition target pattern belongs from among one or more types, based on the gradient feature amount and the sum-difference feature amount calculated from the recognition target pattern and part or all of the learning parameters acquired by the learning unit.

29. The learning identification apparatus of claim 28, wherein:

30. A learning identification method is characterized by comprising the following steps:

extracting a gradient feature quantity including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between the luminance of the input learning object pattern on each coordinate and the luminance of the periphery thereof;

calculating a predetermined sum-difference feature quantity by summing the gradient intensity values conforming to the gradient direction included in a predetermined gradient range indicating a range of the predetermined gradient direction, and subtracting the gradient intensity values conforming to the gradient direction included in other gradient ranges adjacent to the predetermined gradient range from the calculated sum value, based on the extracted gradient feature quantity;

acquiring a learning parameter on each coordinate based on a predetermined learning algorithm by using the gradient feature quantity and the sum-difference feature quantity;

calculating a predetermined sum-difference feature quantity by summing gradient intensity values conforming to the gradient directions included in a predetermined gradient range based on the extracted gradient feature quantity, and subtracting the gradient intensity values conforming to the gradient directions included in the other gradient ranges adjacent to the predetermined gradient range from the calculated sum value;

the type to which the recognition target pattern belongs is recognized from one or more types based on the gradient feature amount and the sum-difference feature amount calculated from the recognition target pattern and part or all of the learning parameters acquired by the learning unit.

31. The learning identification method of claim 30, wherein:

32. A recording medium storing a learning identification program, characterized by causing:

the computer works as follows:

the gradient feature extraction unit extracts a gradient feature amount including a gradient direction and a gradient intensity value thereof on each coordinate based on a variation amount between the brightness of the input recognition target pattern on each coordinate and the brightness of the periphery thereof;

the sum-difference feature extraction unit that calculates a predetermined sum-difference feature amount by summing the gradient intensity values that conform to gradient directions included in a predetermined gradient range, based on the extracted gradient feature amount, and subtracting the gradient intensity values that conform to the gradient directions included in the other gradient range adjacent to the predetermined gradient range from the calculated sum;

the identification unit identifies a type to which the identification target pattern belongs from one or more types based on the gradient feature amount and the sum-difference feature amount calculated from the identification target pattern and part or all of the learning parameters acquired by the learning unit.

33. The recording medium storing the learning identification program of claim 32, wherein: