JP2009151350A - Image correction method and device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a technique for finding efficiently a correction amount according to the preference of the human being who judges the correction precision of an image. <P>SOLUTION: A correction device is provided with a means 5 for calculating a feature amount of an input image, means 2, 4 for assigning a classification attribute and the correction amount about the input image, a means 7 for calculating a classification coefficient for classifying the input image, by a learning method of a neural network using the feature amount and the assigned classification attribute, a means 8 for classifying the input image into a category by the learning processing of the neural network using the classification coefficient, a means 9 for calculating a correction amount coefficient of the category-classified input image, by the learning processing of the neural network using the feature amount and the assigned correction amount, and a means 11 for estimating the correction amount to be applied to the input image, by the learning processing of the neural network using the correction amount coefficient. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a technique for correcting an image, and more particularly to a technique that uses a learning process of a neural network for calculating the correction amount.

  The correction amount in image correction such as color correction and gradation correction needs to be set to an optimum value according to the image. This optimum value depends on the image content and also on the viewer's preference. Therefore, for ideal image correction, it is desirable for a human to manually adjust the correction amount for the image. However, enormous man-hours are required to manually perform a large amount of image correction.

  As a method for automatically controlling the correction amount in place of the complicated manual image correction, for example, there is an automatic image correction method using histogram analysis as described in Non-Patent Document 1 described later.

  As another method related to automatic correction, a method for calculating an image correction amount using a learning process using a neural network has been proposed in recent years. By providing image input data and teacher data to the neural network, it becomes possible to learn complex phenomena appearing in the image as neuron coefficients. Here, a combination of input data and teacher data is called training data, and calculating a coefficient of a neural network using the input data and teacher data is called learning.

  As a method for learning a coefficient of a neural network, there is a method in which an image feature amount is used as input data and an image correction amount is used as teacher data. For example, Patent Document 1 described below describes a technique in which an image signal is input and a coefficient learned using an operator-supplied correction value as a teacher value is used for image correction. In this technique, highlight value, middle value, shadow value, toner density, humidity, and the like are input to the neural network, and the gradation correction coefficient is learned as the output of the neural network.

  Patent Document 2 described later describes a method of calculating image quality characteristic values from an input image and learning gradation correction parameters using a neural network. This document further describes that additional learning is performed according to an operator instruction, and the network is replaced when the error of the new neural network after the additional learning is smaller than the error of the old network.

  As described above, the methods described in Patent Documents 1 and 2 both learn feature quantities detected from an image using a neural network, and apply the learning result to image correction.

  Here, FIG. 12 schematically shows a configuration of a general apparatus for calculating the image correction amount. The apparatus 100 extracts an amount of features from an input image, an image correction unit 101 that corrects an input image, a correction amount specification unit 102 for an operator to specify a correction amount, a monitor 103 that displays an image, and the like. Feature extraction unit 104, storage unit 105 that accumulates feature amounts and correction amounts, coefficient learning unit 106 that learns the relationship between feature amounts and correction amounts as a coefficient of a neural network, and an input image using the learned coefficients And a correction amount estimation unit 107 for estimating a correction amount for.

An example of the image correction unit 101 is gradation correction such as γ correction. As an example of the correction amount specifying unit 102, there is a method of specifying a γ value with a keyboard. As an example of the feature extraction unit 104, there is a method of creating a histogram of RGB values in an image and extracting the maximum value, minimum value, and average value of each. Examples of the storage unit 105 include an HDD and a nonvolatile memory. As an example of the coefficient learning unit 106, there is a method of executing coefficient calculation processing by an error propagation method in a three-layer neural network.
Japanese Patent Laid-Open No. 09-18716 JP 2006-31440 A A. Inoue and J. Tajima, “Adaptive Quality Improvement Method for Color Images”, Proc. Of SPIE, Vol. 2179, pp.429-439, 1994

  However, as in the methods described in Patent Documents 1 and 2, the method of using only the feature amount obtained from the image for learning the correction amount has the following problems.

  First, it is difficult to improve the correction accuracy of an image by learning using only feature amounts. This is because the subjective sense of human influences the determination of the correction accuracy for the image displayed on the monitor. Therefore, even if only the feature amount quantitatively obtained from the input image is used for correction, it is difficult to increase the correction accuracy depending on human subjectivity.

  On the other hand, to optimize the correction amount coefficient, it is beneficial to increase the number of times the neural network is learned. Therefore, in the case of the above-described technique, it is considered that the accuracy is improved by increasing the number of learnings of the coefficient learning unit 106 (FIG. 12). However, as the learning history is accumulated, the calculation becomes complicated, which causes a problem that a large processing load is applied to the coefficient learning unit 106 and the processing time is prolonged.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide a technique for efficiently obtaining a correction amount according to a human preference for determining the correction accuracy of an image.

  An image correction method according to the present invention calculates a feature amount of an input image, and classifies the input image by a neural network learning process using the feature amount and a classification attribute specified for the input image. A classification coefficient is calculated, the input image is classified into categories by a neural network learning process using the classification coefficient, and a neural network learning process using the feature amount and a correction amount specified for the input image. The correction amount coefficient of the input image classified into the category is calculated, and the correction amount to be applied to the input image is estimated by a neural network learning process using the correction amount coefficient.

  An image correction apparatus according to the present invention comprises: means for calculating a feature amount of an input image; means for specifying a classification attribute and a correction amount for the input image; and a classification attribute specified for the feature amount and the input image. Means for calculating a classification coefficient for classifying the input image by learning processing of the used neural network; means for classifying the input image into categories by learning processing of the neural network using the classification coefficient; Means for calculating a correction amount coefficient of the input image classified into the category by a learning process of a neural network using the amount and a correction amount specified for the input image, and a neural network using the correction amount coefficient Means for estimating a correction amount to be applied to the input image.

  A program according to the present invention causes a computer to function as the image correction apparatus.

  According to the present invention, it is possible to obtain a correction amount commensurate with information specified for an input image. Further, since the correction amount coefficient is learned for each category of the input image, the calculation amount for obtaining one correction amount coefficient can be suppressed. Therefore, it is possible to improve the correction accuracy of the image and increase the efficiency of the correction process.

  FIG. 1 shows the configuration of the first embodiment of the present invention. The image correction apparatus 101 according to the present embodiment includes an image correction unit 1 that performs correction by image processing on an input training image, and a correction amount designation unit 2 for an operator to specify a correction intensity (correction amount). The monitor 3 for displaying the image, the classification attribute specifying unit 4 for the operator to specify the classification attribute of the input image, the feature extracting unit 5 for extracting the feature quantity from the input image, and the extracted feature quantity specified An accumulation unit 6 that accumulates correction amounts and classification attributes, a classification coefficient learning unit 7 that calculates classification coefficients using the classification attributes and feature amounts stored in the accumulation unit 6, and an input image using the classification coefficients as a category An automatic classification unit 8 that performs automatic classification, a category-specific coefficient learning unit 9 that calculates a correction amount coefficient using the feature amount and correction amount of an input image classified into a category, and an input image using a correction amount coefficient And a correction amount estimation unit 11 for estimating the correction amount. That.

  The category-specific coefficient learning unit 9 includes a category N coefficient learning unit 10 that learns correction amount coefficients of the Nth category (N = 1, 2,..., N) in classification.

  Examples of the image correction unit 1 include a gradation correction unit such as γ correction and brightness correction, and a color correction unit such as saturation enhancement. As an example of the correction amount specifying unit 2, there is a method of specifying a saturation emphasis amount, a lightness correction amount, a γ value, and the like from a terminal using a keyboard. Examples of the storage unit 6 include an HDD and a non-volatile memory.

  As an example of the classification attribute specifying unit 4, there is a method in which an operator specifies using a screen as shown in FIG. The illustrated classification attribute designation screen 31 includes an input image display area 32, an attention area designation 33, an illumination designation 34, and a subject designation 35. In the attention area designation 33, it is designated which area of the image is to be corrected. In the lighting designation 34, it is designated under which environment the image is taken. In subject designation 33, the type of subject in the image is designated. These classification attributes (33, 34, 35) reflect physical and psychological factors affecting the image quality after correction.

  Regarding the above classification attribute, first, the attention area designation 33 will be described. It is considered that humans focus on only the region of interest in the target input image. For example, in the case of an image of a landscape with a person in the center and a monotonous background, attention is often focused only on the person. In this case, even if the background color is somewhat bad, there is almost no visual effect, and the image quality of the central person area is a factor that determines the overall quality of the image. By assigning such information about a person's “attention” psychological activity as one of the classification attributes, an optimal image correction amount can be calculated.

  FIG. 3 shows a specific example of the attention area designation 33. The first attention area designation 41 designates the entire screen. This indicates that an instruction is given to correct the entire screen equally. The second attention area designation 42 designates the vicinity of the center of the image. This indicates that an instruction is given to intensively correct the vicinity of the center of the image. The third attention area designation 43 designates the vicinity of the head of the person shown in the image. This indicates that the person area is instructed to be corrected mainly.

  Next, the illumination designation 34 and the subject designation 35 will be described. The correction amount of the image is affected by what is captured in the environment of interest (FIG. 3) and in what environment. The object color observed by a human is calculated by integrating the spectral distribution of illumination, the reflectance of the object, and the human visual sensitivity distribution. If the human visual sensitivity is constant, the object color is considered to be determined by the illumination light and the reflectance of the object. However, photographing with a digital camera further adds CCD spectral characteristics, gamma correction, AWB influence, and the like, and thus there is a difference from human appearance. Therefore, by specifying the type of the illumination and the subject as the classification attribute, a certain restriction is given to the learning of the correction amount, thereby making it possible to easily calculate the optimum value.

  FIG. 4 shows a menu 51 used for specifying the illumination (34). In the illustrated menu 51, there are three types of “environment light”, “object incident light” irradiated on an object, and “state” which is an interaction of light such as backlight and forward light. Each lower layer displays specific options in a menu format. As an example of the illumination designation (34), there is a method in which the operator selects an item corresponding to the input image from the menu 51, and inputs the instruction.

  FIG. 5 shows a menu 61 used for subject designation (35). The illustrated menu 61 has three types of items, “person”, “natural object”, and “artificial object”. In each lower layer, specific object names such as “face”, “flower”, and “car” as options are displayed in a menu format. As an example of the subject designation (35), there is a method in which the operator selects an item corresponding to the input image from the menu 61 and inputs the instruction.

  As an example of the feature extraction unit 5 in FIG. 1, there is a method of creating a histogram of RGB values of an image and extracting the maximum value, minimum value, and average value of each. Further, as shown in FIG. 6, the image may be divided into a plurality of partial areas. That is, a histogram 81 is created from the partial areas of the image, and the RGB maximum, minimum, and average values in each partial area, and the maximum, minimum, and average values of saturation in the partial areas and the luminance in the partial areas are displayed. In this method, the maximum value, minimum value, average value, etc. are extracted and used as feature amounts.

  The classification coefficient learning unit 7 learns a classification coefficient for classifying input images having similar image correction tendencies using a neural network. The actual classification process using this classification coefficient is performed by the automatic classification unit 8 at the subsequent stage. Examples of categories in which the correction tendency seems to be similar to the input image include landscapes, people, artifacts, flowers, and night views. The classification coefficient learning unit 7 assigns a teacher category number to the input image using the item specified by the classification attribute specifying unit 4 before learning the classification coefficient.

  The teacher category number is a number of a category that becomes teacher data when learning the classification coefficient. The teacher category number is assigned according to a predetermined rule using a part or all of the classification attribute designation items (FIGS. 3 to 5).

  As an example of a method of assigning a teacher category number, there is a method of using a selection result for “ambient light” and “state” in the illumination designation menu 51 (FIG. 4). In this case, 14 types (7 × 2 = 14) of categories are formed by “(clear sky, cloudy, shade, night view, fluorescent lamp, tungsten, etc.) × (forward light, backlight)”. Then, any number of “1” to “14” is assigned to the input image. Furthermore, when seven types under the “object incident light” in the same illumination designation menu 51 are added, 98 types (7 × 7 × 2 = 98) categories are formed.

  In the above example, the lighting specification (34) was used to form the category, but in addition to this, the attention area specification (33) and the subject specification (35) were each independently or included the lighting specification. It can also be used in combination. If the number of categories is increased too much, a large amount of training data, that is, a combination of input data and teacher data is required. Therefore, for example, it is practical to use about four types of illumination designation and subject designation.

  FIG. 7 shows a configuration example of categories. The category number list 91 shown in the figure uses a combination of illumination designation (34) and subject designation (35) for forming a category. For example, a teacher category number “3” is assigned to an input image in which “outdoor” is designated as the illumination and “night view” is designated as the subject.

  As a method for realizing the classification coefficient learning unit 7 in FIG. 1, for example, the coefficients of a neural network such as a single-layer perceptron, a multi-layer perceptron, and an RBF network (Radical Basis Function Network) are used. There is a way to learn.

  A single layer perceptron will be described. The single-layer perceptron is a neural network that performs discrimination according to the following equation (1).

式（１）において、「x_i」（i=1,2,…,n）はｎ次元の特徴量を表し、「a_i」（i=1,2, …,n）は係数を表し、「y」は推定量を表す。通常、ニューラルネットワークのモデルでは、係数「a_i」の他にオフセット値「a₀」が必要であるが、値が常に「１」である特徴量を「x_i」に含めることにより、上記の式（１）を、オフセット値を含むものとして取り扱うことができる。また、関数U(x)は、出力関数と呼ばれ、例えば、しきい値ｔ以上の場合は「１」を出力し、ｔ未満の場合には「−１」を出力する関数である。

In Expression (1), “x _i ” (i = 1, 2,..., N) represents an n-dimensional feature quantity, “a _i ” (i = 1, 2,..., N) represents a coefficient, “Y” represents an estimated amount. Normally, a neural network model requires an offset value “a ₀ ” in addition to the coefficient “a _i ”. However, by including a feature value whose value is always “1” in “x _i ” Equation (1) can be treated as including an offset value. The function U (x) is called an output function. For example, the function U (x) outputs “1” when it is equal to or greater than the threshold value t, and outputs “−1” when it is less than t.

In the learning process in the classification coefficient learning unit 7, the teacher category number for the input image is given to “y” in equation (1). “X _i ” is an n-dimensional feature amount of the input image. The learning in the classification coefficient learning unit 7 is to obtain the classification coefficient “a _i ” based on a combination of “x _i ” (feature amount) and “y” (teacher category number) of a plurality of input images. Point to.

  Here, when the function U (x) is considered as “U (x) = x”, Expression (1) is a linear discriminant function. This can be expressed in matrix form as shown in equation (2).

式（2）において、行列「X」は、複数の入力画像から得られた特徴量「x_i」の集合を表す。行列「Y」は、それら入力画像に対応する「y」の集合を表す。ベクトル「a」は、係数「a_i」を表す。これを解くと、次の式（3）の行列演算によって係数「a_i」を求めることができる。

In Equation (2), the matrix “X” represents a set of feature quantities “x _i ” obtained from a plurality of input images. The matrix “Y” represents a set of “y” corresponding to the input images. The vector “a” represents the coefficient “a _i ”. When this is solved, the coefficient “a _i ” can be obtained by the matrix operation of the following equation (3).

式（３）により求められた係数「a」を用いて式（３）の演算を行うことで、後述の画像分類処理や、補正量の推定を行うことができる。

By performing the calculation of Expression (3) using the coefficient “a” obtained by Expression (3), it is possible to perform image classification processing, which will be described later, and estimation of the correction amount.

A multilayer perceptron will be described. FIG. 8 shows a three-layer neural network 71 that is a three-layer multilayer perceptron. The three-layer neural network 71 is a neural network having a feature amount “x _i ” as an input layer, a classification category as an output layer, and one layer as an intermediate layer. The three-layer neural network can be expressed by the following equation (4).

式（４）において、「x_i」（i=1,2,…,n）はｎ次元の特徴量を表し、「a_i」（i=1,2,…,n）及び「b_j」（j=1,2,…,m）は係数を表し、「y」は推定量を表す。関数f(x)としては、次の式（5）のシグモイド（Sigmoid）関数を利用することができる。

In Expression (4), “x _i ” (i = 1, 2,..., N) represents an n-dimensional feature quantity, and “a _i ” (i = 1, 2,..., N) and “b _j ” (J = 1, 2,..., M) represents a coefficient, and “y” represents an estimated amount. As the function f (x), the sigmoid function of the following equation (5) can be used.

なお、３層ニューラルネットワーク71のような多層パーセプトロンは、一般に、バックプロパゲーション法（Backpropergation：誤差逆伝播法）という逐次計算によって係数学習を行うことができる。

In general, a multilayer perceptron such as the three-layer neural network 71 can perform coefficient learning by a sequential calculation called a back propagation method (Backpropergation).

  The RBF network will be described. The RBF network can be expressed by the following equation (6).

式（６）において、「x_i」（i=1,2,…,n）はｎ次元の特徴量を表し、「c_i」（i=1,2,…,m）は係数を表し、「z_j」（j=1,2,…,m）は中間ユニットを表す。関数「φ」には、式（７）が利用される。

In Expression (6), “x _i ” (i = 1, 2,..., N) represents an n-dimensional feature quantity, “c _i ” (i = 1, 2,..., M) represents a coefficient, “Z _j ” (j = 1, 2,..., M) represents an intermediate unit. Expression (7) is used for the function “φ”.

RBFネットワークは、中間ユニット「z_j」として、訓練サンプル近傍の値を採用することができる。従って、「z_j」を訓練サンプルの分布に応じて予め決めておくことができ、極端な場合は全ての入力特徴量を「z_j」として用いることもできるので、実際に推定するのは、係数「c_j」（j=1,2,…,m）だけである。

The RBF network can employ values near the training sample as the intermediate unit “z _j ”. Therefore, “z _j ” can be determined in advance according to the distribution of the training sample, and in the extreme case, all input feature values can be used as “z _j ”. Only the coefficient “c _j ” (j = 1, 2,..., M).

Further, in the equation (6), when the function U (x) is considered as “U (x) = x”, the coefficient “c _j ” can be obtained by the equation (8).

In equation (8), the matrix “S” represents the set of intermediate unit outputs “s _j ”. The matrix “Y” represents a set of estimators “y” corresponding to the samples. The vector “c” represents the coefficient “c _j ” (j = 1, 2,..., M). When this is solved, the coefficient “c _j ” can be obtained by the matrix operation of the following equation (9).

なお、分類係数学習部7における学習において、オペレータから指定される分類属性（図２）の情報を、入力画像の特徴量に対する重み付けに利用することができる。例えば、分類属性のうちの注目領域に対しては、照明や被写体といった他の属性の２倍の重みをつけて学習するというものである。

In the learning by the classification coefficient learning unit 7, the classification attribute information (FIG. 2) specified by the operator can be used for weighting the feature amount of the input image. For example, the attention area of the classification attribute is learned with a weight twice that of other attributes such as illumination and subject.

  The automatic classification unit 8 in FIG. 1 classifies input images into categories by a neural network learning process using the classification coefficient obtained by the classification coefficient learning unit 7. Examples of the neural network in the automatic classification unit 8 include a single layer perceptron, a multilayer perceptron, and an RBF network. As the neural network applied to the automatic classification unit 8, the same type as the classification coefficient learning unit 7 is used. The calculation of automatic classification by these three types of neural networks can be performed by the above-described equation (1), equation (4), or equation (6) using the feature amount of the input image.

  Note that the classification result of the input image by the automatic classification unit 8, that is, the category number given to the input image does not necessarily match the teacher category number (FIG. 7) based on the designated classification attribute. This is because the classification coefficient used for classification learning is a value obtained from a mathematical model (neural network) such as Equation (1).

The category-specific coefficient learning unit 9 in FIG. 1 learns the correction amount by the category N coefficient learning unit 10 for each classified category, and calculates the optimal correction amount coefficient for each category. As an example of the learning method in the category-specific coefficient learning unit 9, the neural network used in the classification coefficient learning unit 7 can be used. In the learning process executed by the category-specific coefficient learning unit 9, “x _i ” is an n-dimensional feature amount of the input image, and “y” is a correction amount designated by the operator for the input image.

  The correction amount estimation unit 11 estimates an optimal correction amount to be applied to the input image by a learning process using the correction amount coefficient obtained by the category-specific coefficient learning unit 9. The neural network in the classification coefficient learning unit 7 can also be used for the learning process of the correction amount estimation unit 11. By giving the correction amount obtained by the correction amount estimation unit 11 to the image correction unit 1, an image of the correction result is displayed on the monitor 3. The operator can check the correction accuracy by viewing the image on the monitor 3.

  A series of operations of the image correction apparatus 101 configured as described above will be described with reference to a flowchart shown in FIG. First, when an image to be corrected is input (step S1), the input image is supplied to the image correction unit 1 and the feature extraction unit 5. The feature extraction unit 5 calculates the feature amount of the input image, and stores the result together with the image ID in the storage unit 6 (step S2).

  On the other hand, when the image correction unit 1 outputs the input image to the monitor 3, the operator who has viewed the input image specifies a desired correction amount and also specifies a classification attribute on the screen as shown in FIG. The designated information is input by the correction amount designation unit 2 and the classification attribute designation unit 4 and stored in the storage unit 6 (step S3). In the storage unit 6, the stored feature amounts and classification attributes are stored together with the identification information (ID) of the corresponding input image.

  The classification coefficient learning unit 7 reads the feature amount and the classification attribute of the input image from the storage unit 6, and determines the teacher category number corresponding to the read classification attribute based on the prescribed information such as the list 91 (FIG. 7). To do. Then, the classification coefficient is calculated by giving the determined teacher category number and the read feature quantity to the neural network (step S4).

  The automatic classification unit 8 classifies the input image into categories by giving the neural network the classification coefficient calculated by the classification coefficient learning unit 7 (step S5). Thereby, a category number is assigned to the input image.

  When the category-specific coefficient learning unit 9 recognizes the feature amount of the input image, the correction amount specified by the operator, and the category number calculated by the automatic classification unit 8, the category N coefficient corresponding to the category number (N) The learning unit 10 calculates a correction amount coefficient using a neural network (step S6).

  The correction amount estimation unit 11 estimates the optimum correction amount to be applied to the input image by giving the correction amount coefficient from the category N coefficient learning unit 10 to the neural network (step S7).

  The image correction unit 1 corrects the corresponding input image using the estimated correction amount (step S8), and supplies the image to the monitor 3. As a result, an image as a correction result is displayed on the monitor 3 (step S9), and the operator confirms the image.

  Thus, in this embodiment, information specified by the operator regarding the input image is reflected in the learning process of the neural network. Thereby, the correction amount suitable for the information designated by the operator can be obtained. In addition, since the input image is classified into categories with similar correction tendencies and the correction amount coefficient is learned for each category, the distribution shape to be learned is simplified. Accordingly, the amount of calculation for obtaining one correction amount coefficient can be suppressed, so that the image correction processing can be made more efficient.

  Next, another embodiment of the present invention will be described in detail with reference to the drawings. FIG. 10 shows a configuration of the image correction apparatus 102 in the present embodiment. In the present embodiment, a structural difference from the above-described embodiment (FIG. 1) is that a correction amount distribution analysis unit 12 is added. The correction amount distribution analysis unit 12 analyzes the distribution from the correction amount obtained from the correction amount estimation unit 11 for each category, and determines whether or not the categories need to be integrated or separated based on the analysis result.

  The operation of this embodiment will be described with reference to the flowchart shown in FIG. Here, the difference from the operation (FIG. 9) in the above-described embodiment will be mainly described.

  The correction amount distribution analysis unit 12 obtains an average value for each category of the correction amounts obtained from the correction amount estimation unit 11, and compares them between the categories (step S11). As a result of the comparison, when a combination of categories that approximate the average value of the correction amounts is detected (step S12: Yes), it is determined that the categories of the combination should be integrated (step S13). Note that the number of categories to be integrated is not limited to two, and may be three or more.

  Further, the correction amount distribution analysis unit 12 obtains the distribution width of the correction amount in each category, and compares them with the specified distribution width (step S14). As a result of the comparison, if a category showing a distribution width exceeding the specified width is detected (step S15: Yes), it is determined that the category should be divided into a plurality of categories (step S16). The number of divisions is not limited to a fixed value such as two divisions, and may vary according to the calculated distribution width. In the latter case, a threshold for fluctuation is set in advance.

  On the other hand, if a combination of categories that approximate the average value of the correction amount is not detected (step S12: No), and a category with a distribution width that exceeds the specified width is not detected (step S15: No), the categories are currently combined or divided. Is determined to be unnecessary (step S17).

  When category integration or division occurs, the correction amount distribution analysis unit 12 instructs the classification coefficient learning unit 7 to integrate or divide teacher category numbers. Also, a new teacher category number to be assigned to the input image is calculated (step S18), and designated to the classification coefficient learning unit 7.

  As a new teacher category number calculation method, for example, a generally known cluster analysis method such as a K-means method can be used. According to the K-means method, it is possible to automatically classify into K clusters from the correction amount distribution. When this K-means method is used, the correction amount distribution analysis unit 12 instructs the classification coefficient learning unit 7 as a new teacher category number using the cluster number obtained by the cluster analysis.

  Upon receiving the above instruction from the correction amount distribution analysis unit 12, the classification coefficient learning unit 7 gives a new teacher category number to the input image. Then, using the new teacher category number, a new classification coefficient is calculated by the same learning process (FIG. 9: S4) as in the previous embodiment (step S19).

  The automatic classification unit 8 uses the new classification coefficient from the classification coefficient learning unit 7 to categorize the input image by the learning process (FIG. 9: S5) similar to the above-described embodiment (step S20). When the classification coefficient is changed, the classification result of the input image also changes. For this reason, the category-specific coefficient learning unit 9 uses the new classification result (category number N) from the automatic classification unit 8 to calculate the correction amount coefficient by a learning process (FIG. 9: S6) similar to the above-described embodiment. Calculate (step S21). In this way, the classification coefficient learning and the correction amount coefficient learning are repeated until the correction amount distribution analysis unit 12 determines that the integration or division of the categories is unnecessary (step S17).

  According to the present embodiment, since the category used for classification can be corrected using the learning result of the estimated correction amount, the accuracy of the correction amount estimation can be further improved.

  The present invention is not limited to the above embodiments, and may be implemented as a computer program corresponding to these operations or a recording medium storing the program.

It is a block diagram which shows the structure of embodiment of this invention. It is explanatory drawing regarding designation | designated of the classification attribute which concerns on embodiment of this invention. It is explanatory drawing regarding attention area specification which concerns on embodiment of this invention. It is explanatory drawing regarding the illumination designation | designated which concerns on embodiment of this invention. It is explanatory drawing regarding subject specification which concerns on embodiment of this invention. It is explanatory drawing regarding extraction of the image feature-value which concerns on embodiment of this invention. It is explanatory drawing regarding provision of the teacher category which concerns on embodiment of this invention. It is explanatory drawing regarding the 3 layer neural network which concerns on embodiment of this invention. It is a flowchart regarding operation | movement of embodiment of this invention. It is a block diagram which shows the structure of other embodiment of this invention. It is a flowchart regarding operation | movement of other embodiment of this invention. It is a block diagram which shows the structural example of a general image correction apparatus.

Explanation of symbols

101,102 Image correction device
1 Image correction section
2 Correction amount specification section
3 Monitor
4 Classification attribute specification part
5 Feature extraction unit
6 Accumulator
7 Classification coefficient learning unit
8 Automatic classification part
9 Coefficient learning section by category
10 Category N coefficient learning section
11 Correction amount estimation unit
12 Correction amount distribution analysis unit

Claims (11)

Calculate the feature value of the input image,
By a neural network learning process using the feature amount and the classification attribute specified for the input image, a classification coefficient for classifying the input image is calculated,
The input image is classified into categories by a neural network learning process using the classification coefficient,
A neural network learning process using the feature amount and the correction amount specified for the input image calculates a correction amount coefficient of the input image classified into the category,
An image correction method, wherein a correction amount to be applied to the input image is estimated by a neural network learning process using the correction amount coefficient.   The image correction method according to claim 1, further comprising: comparing an average value in a category related to the estimated correction amount between a plurality of categories, and integrating a plurality of categories that approximate the average value.   The image correction method according to claim 1, further comprising: calculating a distribution width within a category related to the estimated correction amount, and dividing a category in which the distribution width exceeds a specified width into a plurality of categories.   The image according to any one of claims 1 to 3, wherein the classification attribute is information specified for some or all of a region of interest, an illumination type, and a subject type for the input image. Correction method.   5. The image correction method according to claim 1, wherein each of the neural networks is any one of a single-layer perceptron, a multi-layer perceptron, and an RBF perceptron. Means for calculating a feature amount of the input image;
Means for specifying a classification attribute and a correction amount for the input image;
Means for calculating a classification coefficient for classifying the input image by a learning process of a neural network using the feature quantity and a classification attribute designated for the input image;
Means for classifying the input image into categories by a neural network learning process using the classification coefficient;
Means for calculating a correction amount coefficient of the input image classified into the category by a learning process of a neural network using the feature amount and a correction amount specified for the input image;
An image correction apparatus comprising: means for estimating a correction amount to be applied to the input image by a neural network learning process using the correction amount coefficient.   7. The image correction apparatus according to claim 6, further comprising means for comparing an average value in a category relating to the estimated correction amount between a plurality of categories and integrating a plurality of categories whose average values approximate.   8. The image correction apparatus according to claim 6, further comprising means for calculating a distribution width within a category relating to the estimated correction amount and dividing the category having the distribution width exceeding a specified width into a plurality of categories. .   The image according to any one of claims 6 to 8, wherein the classification attribute is information specified for some or all of a region of interest, an illumination type, and a subject type for the input image. Correction device.   The image correction apparatus according to claim 6, wherein each of the neural networks is any one of a single layer perceptron, a multilayer perceptron, and an RBF perceptron.   A program for causing a computer to function as the image correction apparatus according to any one of claims 6 to 10.

Images