CN1141665C - Methods of Feature Extraction and Recognition of Microscopic Image - Google Patents

Abstract

The invention belongs to the technical field of image processing, and is a method for feature extraction and recognition of microscopic images, including two parts: extraction of microscopic image features and feature recognition. Among them: the feature extraction of the microscopic image includes microscopic magnification of the image, digitizing the simulated microscopic image, processing it with binarization, adopting normalized coordinates to represent the position information of the edge of the digitized image, and reconstructing the information Processing and template preservation; microscopic image recognition includes microscopic magnification of the image to be recognized, digitization of the microscopic image, information extraction of the edge position of the digitized image, and comparison of the information with the stored feature template. Among them: the criterion of the maximum error limit is used to judge the authenticity of the image recognition. It has the characteristics of simple implementation, clear edge outline and reliable identification, and is especially suitable for the identification and anti-counterfeiting of the authenticity of valuable securities and documents.

Description

Methods of Feature Extraction and Recognition of Microscopic Image

technical field

The invention belongs to the technical field of image processing, in particular to feature extraction and feature recognition of microscopic images.

Background technique

Microscopic images are widely used in remote sensing, research on fractal characteristics of reservoirs, research on microscopic profile characteristics on the surface of optical components, microscopic characteristic analysis of nano-silicon thin film interface structure, and material detection. For different microscopic image features and application purposes, the methods of feature extraction and recognition are different, and there is no unified model. Taking a fingerprint recognition method based on structural features as an example, the extraction and recognition of image features are illustrated. The feature extraction of this fingerprint identification method includes the following steps: collecting fingerprint images and converting them into digital images for input into the computer, binarization based on edge search, smoothing the binarized images, thinning, extracting fingerprint features (including edge , bifurcation, mouth shape, thorn, cross, bridge and other characteristics). Fingerprint matching identification includes extracting the features of the fingerprint to be tested, comparing with the saved feature data, and outputting the identification results. The influence of angle rotation and distortion etc. should be considered in recognition.

Common methods for image edge detection include: differential operator method; template matching method; boundary and curve enhancement techniques, such as iterative boundary detection, relaxation method boundary enhancement, or using parameter space, edge pixels according to boundary or curve parameters Do aggregation; continuous wavelet edge detection; edge focusing; texture edge detection; neural network edge detection, etc.

The differential operator method is to use different forms of differential operators to calculate the spatial derivative of the gray level of each pixel to give a differential sharpening image. Such as the gradient operator:

G(i, j) = Δ _x f(i, j) ² + Δ _y f(i, j) ² (1) Laplacian operator:

G(i,j)=Δ ² _x f(i,j)+Δ ² _y f(i,j) (2)

= f(i+1, j)+f(i-1, j)+f(i, j+1)-f(i, j-1) where f(i, j) is point (i, j) gray value at . The differential operator method is mainly aimed at gray-scale images, highlighting the edge features. Although the edges of the processed image are enhanced, the gray levels of each edge are still not uniform, and the interference of the background still exists, making it difficult to automatically identify the boundary. It is difficult to determine the position of the boundary by the differential operator method. Although the processed image looks obvious to the human eye, it is difficult for the computer to automatically identify the boundary.

The template matching method uses the ideal small edge area to form an edge template, detects the matching degree of each pixel in the image and the template, and gives an edge image. For example, the direction template can detect edges in different directions. But for edge images with random features, it is impossible to make templates in advance.

The boundary and curve enhancement technology uses the neighborhood information of the pixel to enhance the pixel on the basis of edge detection, or uses the parameter space to aggregate the edge pixels according to the boundary or curve parameters, so that the processed image has the same differential operator processing For similar features, the edge position is difficult to automatically identify. Continuous wavelet edge detection and edge focusing also have similar weaknesses.

Textured edge detection requires objects and backgrounds to have different textures, and thus is not suitable for direct application to random images. As for the neural network edge detection method, although there are many algorithms that can be converted into neural network implementations, they do not reflect the essence of the neural network system. Further research is needed to truly construct a feature detection method that imitates the biological visual system. So far there is no ready-made effective method.

Contents of the invention

The purpose of the present invention is to propose a method for feature extraction and recognition of microscopic images according to the microscopic random characteristics of images, which has the characteristics of simple implementation, clear edge outline and reliable recognition, and is suitable for feature extraction and recognition of microscopic images. It can also be applied to common computer image recognition, especially for the identification and anti-counterfeiting of the authenticity of valuable securities and documents.

The present invention proposes a method for feature extraction of a microscopic image, which is characterized in that: comprising the following steps;

1) Enlarge a specific image to become a microscopic simulation image,

2) digitize the microscopic image to become a digitized image,

3) Binarize the digitized image into a binary image,

4) Carry out the feature extraction of edge position to this binary image,

5) and normalize the feature and save the feature template;

The present invention proposes a method for feature recognition of microscopic images, which is characterized in that: comprising the following steps:

1) The image to be recognized is microscopically enlarged to become a microscopic simulation image,

2) digitize the microscopic image to become a digitized image,

3) Binarize the digitized image into a binary image,

4) Carry out the feature extraction of edge position to this binary image,

5) Normalize the feature,

6) and comparing the feature with the stored feature template to judge the authenticity of the image to be recognized.

Principle and characteristic of the present invention:

The main contribution of the present invention is to discover the microcosmic random feature characteristic of the image, and use this characteristic as the basis of the feature extraction and feature recognition method of the present invention, making the method simple and reliable.

The image involved in the present invention refers to all lines, dots, characters and images formed on an image carrier (such as paper) by means of reproduction or non-replication. It is found through research that the edge or internal features of the image, especially the microscopic edge or internal features of the image present a random property. As shown in Figure 1, Figure 1(a) is a macroscopic view of two identical straight lines 1 and 2 with a width of 0.3 mm, and Figure 1(b) is a microscopic enlarged view of the left ends 11 and 21 of the two straight lines, as can be seen , the shapes of the end contours of the two graphs are completely different in the microscopic state, and the edge correlation between them is very low. It can also be seen that the distribution of image edge points is random and there is no regularity at all. In addition, these lines are continuous macroscopically, but not microscopically. There are many discontinuous areas in the lines, and their distribution is also random.

Since the distribution of these edge points and internal discontinuous areas are random, any two such images have a low possibility of overlapping edges and a low probability of having the same internal areas. The law of random distribution of edge points of this image is unique, and any two images are different. The randomness and uniqueness of the edge profile of the microscopic image can be used as the feature of the image, and the present invention uses this feature of the image as the basis of the method for image recognition.

The image feature extraction method of the present invention mainly includes the detection and feature extraction of image edge contours and the generation and preservation of feature templates. The feature recognition then first adopts the same steps as the feature extraction process to extract the features of the recognized image, and then The feature is compared with the saved feature template, and the recognition result is given.

Description of drawings

Figure 1 is a microscopic image of two identical straight lines and their endpoints; where (a) is a macroscopic image, and (b) is a local microscopic enlarged image.

Fig. 2 is a flow chart of image feature extraction in the present invention.

Fig. 3 is a flow chart of image feature recognition in the present invention.

Fig. 4 is the result of binarization processing of the microscopic image in Fig. 1 .

Fig. 5 shows the extraction method and result of edge position information.

Fig. 6 is an explanatory diagram of a rotated image.

Detailed ways

The content and realization principle of the present invention are described in detail below through a preferred embodiment of the present invention in conjunction with the accompanying drawings:

The method for the feature extraction and identification of a kind of microscopic image that the present invention proposes, it comprises the extraction of microscopic image feature and the identification two parts of feature, as shown in Figure 3, wherein, image feature extraction comprises the following steps:

1) Enlarge a specific image to become a microscopic simulation image,

2) digitize the microscopic image to become a digitized image,

3) Binarize the digitized image into a binary image,

4) Carry out the feature extraction of edge position to this binary image,

5) and normalize the feature and save the feature template;

Image feature recognition includes the following steps:

6) The image to be recognized is microscopically enlarged to become a microscopic simulation image,

7) digitize the microscopic image to become a digitized image,

8) Binarize the digitized image into a binary image,

9) Carry out the feature extraction of edge position to this binarized image,

10) Read the sample data;

11) directly compare and recognize the position information of the image to be recognized with the template data, and end if the recognition result is "true";

12) If the recognition result is "false", the position information of the image to be recognized is compared with the template data, and the recognition result is "true" and then ends;

13) If the recognition result is "false", the position information of the image to be recognized is rotated and compared with the template data to obtain the recognition result;

14) Output the recognition result.

The steps of feature extraction and feature recognition in the present invention will be described in detail below by comparing the two straight lines in FIG. 1 as an embodiment.

1. Feature extraction

The straight line 1 is enlarged into a microscopic image through the optical system, and the microscopic image is collected by an image acquisition device to become a grayscale digital image 11; the grayscale digital image is input into the computer, and it is binary valued according to a predetermined threshold Transform the grayscale digital image into a black and white binary (0 and 1) image.

The present invention proposes a threshold selection method, as shown in formula (3),

p _T ＝p _min +α(p _max -p _min ) (3) where p _T is the selected threshold; p _min is the minimum gray value in the image; p _max is the maximum gray value; α is a constant, which can be determined according to the experiment The result is OK. The principle of α selection is: the binarized image can separate the useful information from the background without losing edge information. Experiment with a certain type of image, change the size of α, and select an appropriate value according to the result of binarization to meet the above requirements. The threshold value determined by this method has the characteristic of automatically adjusting the threshold value with the change of light intensity.

Fig. 4 is the image after binary processing of Fig. 1(b) using formula (3). It can be seen that the edge information of the two linear images 13 and 23 after binarization is obvious, and the edge correlation between the two is very low .

After the image has been binarized, the random feature information of the edge position can be extracted from the binarized image. The specific method is: search from left to right along each line of the binarized image, stop searching when encountering a point with a gray value of 0, and use this point as an edge point, and record its position information x _i (top left of the image The angle is the origin of the coordinate axis, the x-axis is positive to the right, and the y-axis is positive downward), as the extracted position information; this method actually regards the leftmost black point of each row as an edge point.

For different lines, due to their different widths, the number of edge points is also different. If the position information of all points on the left edge is extracted and saved, the number of saved data will be different, which is not easy to compare and unify, and to find the upper and lower edges The starting point is more difficult. Therefore, in this preferred embodiment, the information in the middle part of the left edge is intercepted and saved, that is, the edge information whose ordinate is between y _a and y _b is extracted. As shown in Figure 5(a), the vertical coordinates of the positions of the straight lines A and B are y _a and y _b respectively, from the line y _a to the line y _b , search from left to right respectively, and obtain the edge position of each line Coordinates x _a , . . . , x _i , . . . x _b . Fig. 5(b) is an edge contour graph drawn according to coordinate values x _a , ..., x _i , ... x _b .

Then normalize the above coordinate values, the method is: compare the size of these edge point position coordinates, find out the minimum value x _n among them, and subtract this minimum value from each coordinate value x _i to get new data x _a -x _n , ..., x _i -x _n , ... x _b -x _n , save these new data as feature templates for recognition.

Second, feature recognition,

When identifying, first use the same steps as the above feature extraction to extract the random feature information of the image to be tested, and then compare it with the feature template saved in advance.

Comparison recognition is actually to compare whether the edge point positions of the current image and the saved original image are consistent. If they match within the allowable range of error, the recognition result is "true", otherwise it is "false".

When comparing and recognizing, due to the influence of the front and rear environments, all data cannot be completely overlapped, so a maximum error limit should be determined according to specific requirements and experimental results. As long as the recognition result meets this requirement, the recognition is passed, otherwise the recognition fails. This maximum error limit should meet two requirements: first, no misjudgment can occur, and the false image is recognized as true; second, the recognition rate is high, that is, the probability of recognizing a true image as false is very low or even zero.

The identification process of this embodiment is described in detail as follows:

In FIG. 5 , the y-axis coordinate values of A and B are y _a and y _b respectively, so the total number of collected edge points is (y _b −y _a +1).

When performing image recognition, first follow the same steps as image feature extraction to extract the left edge features (y _b -y _a +1 feature data) of the sample line 2 to be recognized, and then perform normalized coordinate processing, which can avoid sample The position effect brought by the left and right movement. Then read the saved template data, and then compare and identify according to the maximum error limit criterion (Formula (4) and Formula (5)).

Although the shape of the newly obtained image edge is similar to the original edge, due to the influence of acquisition conditions, it is impossible to completely coincide with the original edge. Therefore, when identifying, a maximum deviation value x _i is specified, and each edge point only needs to satisfy the formula ( 4), it is considered that the edges are coincident.

|x _s -x _m | _≤xi (4) where x _s is the saved feature template data; x _m is the currently detected data; x _i is the maximum deviation allowed. The principle selected by _xi should be able to prevent misjudgment, but also to be able to identify the real object. In a preferred embodiment, the value of x _i is 3. According to the quality of the image and a large number of experimental results, _xi can fluctuate slightly. The key is to improve the recognition rate as much as possible on the premise of preventing misjudgment.

As long as the final comparison result satisfies formula (5), the edges are considered to be coincident, and the recognition result is "true".

N _c ≥N _i ×p (5) where N _c is the total number of edge points satisfying formula (4); N _i is the total number of data (y _b -y _a +1); p is the accuracy rate that needs to be satisfied , in a preferred embodiment, the value of p is 90%. The larger the value of p, the lower the misjudgment rate, but it will reduce the recognition rate; the smaller the value of p, the higher the recognition rate, but there may be misjudgment. The p-value of a certain type of sample should be determined based on a large number of experiments, and the recognition rate should be improved without misjudgment.

During image feature extraction and feature recognition, there will always be up, down, left, right, and angle deviations in the position of the sample, so the image needs to be rotated and translated.

If the result of the direct comparison is "true", the comparison process ends; if the result of the direct comparison is "false", a translational comparison is performed.

The position information of y coordinates from y _an to y _b+n , a total of y _b -y _a +2n+1 edge points is detected x _an ,..., x _a ,..., x _b ..., x _b+n , where n is an integer, is the amount to move up or down. When comparing, select y _b -y _a +1 values in turn from them, and compare them with the saved feature template. In this way, there are 2n+1 sets of data in total, and they are: x _an ～x _bn , x _a-n+1 ～x _b-n+1 , ..., x _a ～x _b , ..., x _{a+ n-1} ～x _b+n-1 and x _a+n ～x _b+n , each set of data is regarded as y-coordinates from y _a to y _b , and compared with the characteristic template data. As long as one group of data satisfies formula (5), the edges are considered to be overlapped, and the recognition result is "true". Translational comparison actually includes the process of direct comparison.

If the result of the translation comparison is "true", the comparison process ends; otherwise, the rotation comparison is performed.

Rotation comparison is to first rotate the whole image by an angle, and then perform translation comparison.

As shown in Figure 6, the coordinates of point A and point B in the same coordinate system are (x, y) and (x ₀ , y ₀ ) respectively, and OA rotates an angle θ relative to OB, then the coordinates of the two are as follows Relational formula:

x＝x ₀ cosθ-y ₀ sinθ (6)

y＝x ₀ sinθ+y ₀ cosθ

When rotating the image, first determine an angle step size α, then rotate the entire image by the angle α or -α according to formula (6), and then perform a translation comparison. If the comparison result is "true", the comparison ends; if the recognition result is "Pseudo", and then rotate the original image to 2α and -2α for translation comparison. If the comparison result is "true", end the comparison; if the recognition result is "false", then rotate the original image by 3α and -3α,...nα and -nα, and perform translation comparison respectively. If the recognition is still not successful when the rotation angles are nα and -nα, the recognition result "false" is output, and the recognition process ends.

Claims (7)

1. A method for feature extraction of a microscopic image, characterized in that: comprising the following steps; 1) Enlarge a specific image to become a microscopic simulation image, 2) digitize the microscopic image to become a digitized image, 3) Binarize the digitized image into a binary image, 4) Carry out the feature extraction of edge position to this binary image, 5) and normalize the feature and save the feature template. 2. The method for extracting microscopic image features as claimed in claim 1, characterized in that: the method for feature extraction of said edge position specifically comprises: searching from left to right along each line of the binarized image, when encountering Stop searching when it reaches a point with a gray value of 0, and use this point as an edge point, and record the coordinate information x _i of its position as the extracted position information; select the position information of the set number of lines as the feature template of a specific image , where the upper left corner of the image is the origin of the coordinate axes, the x-axis is positive to the right, and the y-axis is positive to the downward. 3. The method for extracting microscopic image features as claimed in claim 1, characterized in that: said normalization processing method is: extracting position coordinate information x _i of a specific image with a set number of rows, and comparing all coordinate information The size of x _i , find the minimum value x _n , subtract this minimum value from each coordinate value x _i , get new data x _i -x _n and save these new data as the specific image during recognition Feature boilerplate. 4. A method for feature recognition of microscopic images, characterized in that: comprising the following steps: 1) The image to be recognized is microscopically enlarged to become a microscopic simulation image, 2) digitize the microscopic image to become a digitized image, 3) Binarize the digitized image into a binary image, 4) Carry out the feature extraction of edge position to this binary image, 5) Normalize the feature, 6) and comparing the feature with the stored feature template to judge the authenticity of the image to be recognized. 5. The method for feature recognition of microscopic images as claimed in claim 4, characterized in that: the method for feature extraction of said edge position specifically comprises: searching from left to right along each line of the binarized image, when Stop searching when encountering a point with a gray value of 0, and use this point as an edge point, record the coordinate information x _i of its position as the extracted position information; select the position information of the set number of lines as the image to be recognized Features, where the upper left corner of the image is the origin of the coordinate axis, the x-axis is positive to the right, and the y-axis is positive to the downward. 6. The method for feature recognition of microscopic images as claimed in claim 4, characterized in that: said normalization processing method is: extract the position coordinate information x _i of the image to be recognized with a set number of rows, and compare all The size of the coordinate information x _i , find the minimum value x _n , subtract this minimum value from each coordinate value x _i , get new data x _i -x _n and save these new data as the image to be recognized Characteristics. 7. The method for identifying microscopic image features as claimed in claim 4, wherein said method of comparing the features of the image to be identified with the stored feature templates comprises the following steps: 1) determine the maximum error limit criterion; 2) First, directly compare the feature of the image to be recognized with the stored feature template, if it satisfies the maximum error limit criterion, and the recognition result is "true", then end; 3) If the recognition result is "false", the features of the image to be recognized are relatively translated, and then compared, and the recognition result is "true", then the end; 4) If the recognition result is "false", the features of the image to be recognized are relatively rotated, and then the translation comparison in step 3) is performed to obtain the recognition result.

Images