CN110188646B - Human ear recognition method based on the fusion of gradient direction histogram and local binary pattern - Google Patents

Abstract

The invention discloses a human ear recognition method based on the fusion of gradient direction histogram and local binary pattern, which solves the problem of low recognition rate of human ear recognition from human ear images. The invention first extracts the gradient direction histogram of the image, reduces the dimension by principal component analysis, then extracts the texture feature of the image by using the local binary mode, then fuses the two features, and finally uses the minimum distance classifier for classification. The present invention improves the recognition rate of human ear recognition through multi-feature fusion, and has good practicability and effectiveness.

Description

Human ear recognition method based on the fusion of gradient direction histogram and local binary pattern

technical field

The invention relates to a human ear recognition method based on the fusion of a gradient direction histogram and a local binary pattern, and belongs to the cross technical fields of biological feature recognition, deep learning, artificial intelligence and the like.

Background technique

As a new biometric identification technology, human ear recognition has received more attention from scholars at home and abroad in the past two years, and has important theoretical significance and practical application value.

Human ear recognition is based on human ear images as the research object for feature recognition, which can be used as a beneficial supplement to other biometric technologies, and can also be used alone in the occasion of individual identification. In biometric-based identification technology, human ear recognition has many advantages, such as the small size of the human ear image, the small amount of calculation, the consistent color distribution of the outer ear, less information loss when converting to a grayscale image, and the human ear is not affected by expressions. The impact of changes, non-intrusive recognition can be achieved, etc.

The current methods of human ear recognition can be divided into two categories according to the extracted features: one is the method based on geometric features. Such methods are easily affected by illumination and imaging angle, and have poor robustness. One is the methods based on algebraic features, such as principal component analysis method, invariant moment III method, wavelet transform method and so on. These methods achieve satisfactory results under the condition that the human ear pose changes little and the image quality is good. However, when the rotation angle of the human ear changes, the two-dimensional image will bring about a large deformation. At this time, the recognition rate of the traditional method will drop sharply. Therefore, a more cost-effective and accurate human ear recognition method needs to be carried out A lot of research work.

SUMMARY OF THE INVENTION

Technical problem: The technical problem to be solved by the present invention is how to use the minimum distance classifier to perform human ear recognition on the input human ear image, so as to improve the training speed and accuracy of human ear recognition.

Technical solution: A human ear recognition method based on the fusion of gradient direction histogram and local binary pattern of the present invention includes the following steps:

Step 1) obtaining the image of the human ear from the human ear image library;

Step 2) Calculate the feature value on each pixel in the human ear image, obtain the gradient direction histogram feature of the human ear image after dividing and standardizing;

Step 3) Use the principal component analysis method to spatially transform the gradient direction histogram of the human ear image, so that the original coordinates are projected to a new space with lower dimensions and mutually orthogonal, so as to realize the gradient direction of the human ear image. Dimensionality reduction of histogram features;

Step 4) calculating the local binary pattern value of each pixel on the human ear image to obtain the local binary pattern feature of the human ear image;

Step 5) cascading the eigenvectors of the gradient direction histogram feature and the local binary pattern feature to obtain a new eigenvector to realize feature fusion;

Step 6) Input to the minimum distance classifier for classification and identification.

in,

Described step 2) is as follows:

Step 21) Perform color standardization on the human ear image obtained in step 1), and uniformly convert the image into a grayscale image. The conversion formula is: H(x,y)=0.3*R(x,y)+0.59* G(x,y)+0.11*B(x,y), where R(x,y), G(x,y), B(x,y) are the red, green, The color value of blue, H(x,y) represents the gray value of each pixel.

Step 22) use Sobel operator to calculate the modulus value and direction angle of each pixel gradient of the image:

Among them, G(x, y) represents the gradient magnitude of the pixel point, α(x, y) represents the gradient direction of the pixel point, and H(x, y) represents the gray value of the pixel point.

Step 23) Space and orientation cells vote with weight. First calculate the direction weights: x(i)=cos(θ), y(i)=sin(θ), θ=θ+π/(N _direction +1), where: i is the direction number; θ is the angle, The initial value is 0; x(i) is the weight of the x-axis difference in the i direction; y(i) is the weight of the y-axis difference in the i direction, and N _direction is the total number of directions, generally set to 9;

Step 24) calculate the amplitude and direction, the amplitude is the mean square value of the image difference of the x or y axis, and the direction value is taken as the weighted maximum value in each direction;

其中B(x)、B(y)分别表示块x轴总数值和y轴总数值，C(x)、C(y)分别代表单元格x轴总数值和y轴总数值，B(size)为块的大小，B(step)是块变化的步长；Step 25) Building blocks and standardizing, summarizing and combining the features in the cells into blocks, the calculation method is:

Among them, B(x) and B(y) represent the total value of the x-axis and the total value of the y-axis of the block respectively, C(x) and C(y) represent the total value of the x-axis and the total value of the y-axis of the cell, and B(size) is the size of the block, and B(step) is the step size of the block change;

Step 26) Summarize the feature values in different directions and blocks, and construct the direction gradient histogram of the image.

Described step 3) is as follows:

Step 31) Calculate the mean value of the gradient direction histogram feature of the corresponding pixel in the human ear image

计算协方差矩阵，其中x_i为需要降维的特征，U^T为协方差矩阵。Step 32) According to

Calculate the covariance matrix, where x _i is the feature that needs to be reduced in dimension, and U ^T is the covariance matrix.

其中y表示主成分特征。p根据实际情况通过实验获得，过大速度会慢，过小会影响准确率。Step 33) Take the first p principal components of the covariance matrix, and perform feature dimension reduction on the eigenvalues of each gradient direction histogram in the human ear image to obtain the gradient direction histogram feature of the dimension reduction by the principal component analysis method of the human ear image, the vector The dimension is p-dimension. The feature dimensionality reduction method is

where y represents the principal component feature. p is obtained through experiments according to the actual situation. If it is too large, the speed will be slow, and if it is too small, the accuracy will be affected.

Described step 4) is as follows:

其中(x_c,y_c)代表3×3邻域的中心元素，它的像素值为i_c，i_p表示邻域内其他像素的值，s(x)是符号函数，当x≥0时为1，否则为0，LBP(x_c,y_c)是中心像素的局部二值模式值的二进制表示。Step 41) Define the central pixel of the 3×3 window as the threshold value, compare the gray values of the remaining 8 pixels with the threshold value in turn, and mark the pixel larger than the central pixel as 1, otherwise mark as 0, and form an 8-bit binary number. The binary representation of the local binary mode value of the window.

where (x _c , y _c ) represents the central element of the 3×3 neighborhood, its pixel value is _ic , _ip represents the value of other pixels in the neighborhood, s(x) is the sign function, and when x≥0, it is 1, otherwise 0, LBP(x _c , y _c ) is the binary representation of the local binary pattern value of the center pixel.

Step 42) Convert the LBP(x _c , y _c ) of each pixel into decimal numbers to obtain the final local binary pattern value, and after summarizing the local binary pattern features of the human ear image.

Described step 6) is as follows:

Input the eigenvectors obtained after the processing of steps 2 to 5 into the minimum distance classifier, respectively, to calculate the eigenvectors of the to-be-recognized human ear image and each known classification The distance between the feature vectors of the human ear images, and the classification corresponding to the smallest distance is the classification of the human ear image to be recognized. The distance calculation formula is as follows:

where d(i) is the distance between the human ear image to be recognized and the i-th known-classified human ear image, t ₁ (i) is the first feature vector of the i-th known-classified human ear image, p ₁ is the first feature vector of the human ear image to be recognized, t _n (i) is the n-th feature vector of the i-th known classified image, and p _n is the n-th feature vector of the image to be recognized.

Beneficial effect: the present invention adopts the above technical scheme compared with the prior art, has the following technical effects:

The invention uses the principal component analysis method to reduce the dimension of the gradient direction histogram feature, filters out a large amount of redundant information, greatly improves the accuracy of human ear recognition, reduces the dimension of the feature vector, and improves the human ear recognition. The speed of recognition; the invention uses the local binary pattern feature and the gradient direction histogram feature to fuse, overcomes some noise interference, increases the robustness of the feature vector, and improves the stability of the human ear recognition algorithm. The present invention has a higher recognition rate for using a single feature for identification; the present invention adopts a minimum distance classifier, which has lower computational complexity and faster speed. Through the application of these methods, the accuracy and stability of human ear recognition are improved, while the computational complexity is reduced, making the system highly cost-effective, specifically:

(1) The present invention adopts multi-feature fusion for classification and identification, which has higher accuracy than a single feature.

(2) The present invention uses the principal component analysis method to reduce the dimension of the gradient direction histogram feature, filters out a large amount of redundant information, and greatly improves the accuracy of human ear recognition.

(3) The present invention uses the local binary pattern feature and the gradient direction histogram feature, overcomes some noise interference, increases the robustness of the feature vector, and improves the stability of the human ear recognition algorithm.

(4) The present invention uses the principal component analysis method to reduce the dimension of the gradient direction histogram feature. Compared with the traditional gradient direction histogram feature, the dimension of the feature vector is reduced, and the speed of the entire human ear recognition is improved.

(5) Compared with other classifiers, the minimum distance classifier based on Euclidean distance adopted in the present invention has lower computational complexity and improves the speed of human ear recognition.

Description of drawings

Figure 1 is the flow of the human action recognition method based on convolutional neural network.

Detailed ways

In a specific implementation, FIG. 1 is a flowchart of a method for recognizing human behavior based on a convolutional neural network.

This example uses the human ear experiment library of University of Science and Technology Beijing as the experimental object, which contains 77 human ear images. There are four human ear images for each human ear in the human ear library, namely: the frontal image of the human ear under normal conditions, the image rotated by +30 degrees and -30 degrees of the human ear, and the frontal image of the human ear under the condition of dim light. .

In the specific implementation, each person has 4 human ear images, of which 3 images are used for training and 1 image is used for testing.

其中B(x)、B(y)分别表示块x轴总数值和y轴总数值，C(x)、C(y)分别代表单元格x轴总数值和y轴总数值，B(size)为块的大小，B(step)是块变化的步长；将不同方向和块上的特征值汇总，构建图像的方向梯度直方图。First, input 3 human ear images of each person into the system, and perform color normalization processing, using H(x,y)=0.3*R(x,y)+0.59*G(x,y)+0.11*B( x,y) converts the image into a grayscale image, where R(x,y), G(x,y), B(x,y) are the red, green, and blue color values of each pixel in the image, respectively , H(x,y) represents the gray value of each pixel; use the Sobel operator to calculate the modulus value and direction angle of the gradient of each pixel of the image; calculate the direction weight, magnitude and direction, the magnitude is the image of the x or y axis The mean square value of the difference, the direction value is taken as the weighted maximum value in each direction; the block is constructed and standardized, and the features in the cells are aggregated and combined into blocks. The calculation method is:

Among them, B(x) and B(y) represent the total value of the x-axis and the total value of the y-axis of the block respectively, C(x) and C(y) represent the total value of the x-axis and the total value of the y-axis of the cell, and B(size) is the size of the block, and B(step) is the step size of the block change; the feature values in different directions and blocks are summarized to construct the directional gradient histogram of the image.

计算协方差矩阵，其中x_i为需要降维的特征，

为人耳图像梯度方向直方图特征的均值，U^T为协方差矩阵。取协方差矩阵前p个主成分，对人耳图像中每一个梯度方向直方图特征值，用

进行特征降维，得到人耳图像的主成分分析法降维的梯度方向直方图特征，其中向量维数为p维，y表示主成分特征。Then, use the principal component analysis method to spatially transform the gradient direction histogram of the human ear image, so that the original coordinates are projected to a new space with lower dimensions and orthogonal to each other, so as to realize the gradient direction histogram of the human ear image. Dimensionality reduction of graph features. The specific method is based on

Calculate the covariance matrix, where x _i is the feature that needs to be reduced in dimension,

is the mean value of the histogram feature of the gradient direction of the human ear image, and U ^T is the covariance matrix. Take the first p principal components of the covariance matrix, and for each eigenvalue of the gradient direction histogram in the human ear image, use

The feature dimension reduction is performed to obtain the gradient direction histogram feature of the principal component analysis method of the human ear image, wherein the vector dimension is p dimension, and y represents the principal component feature.

其中(x_c,y_c)代表3×3邻域的中心元素，它的像素值为i_c，i_p表示邻域内其他像素的值，s(x)是符号函数，当x≥0时为1，否则为0，LBP(x_c,y_c)是中心像素的局部二值模式值的二进制表示，再将其转换成十进制数得到最终的局部二值模式值，得到人耳图像的局部二值模式特征。Then, the central pixel of the 3×3 window is defined as the threshold, and the gray values of the remaining 8 pixels are compared with the threshold in turn. If the pixel is larger than the central pixel, it is marked as 1, otherwise it is marked as 0, and the 8-bit binary number is this window. The binary representation of the local binary mode value of .

where (x _c , y _c ) represents the central element of the 3×3 neighborhood, its pixel value is _ic , _ip represents the value of other pixels in the neighborhood, s(x) is the sign function, and when x≥0, it is 1, otherwise 0, LBP(x _c , y _c ) is the binary representation of the local binary pattern value of the central pixel, and then convert it into a decimal number to obtain the final local binary pattern value, and obtain the local binary pattern of the human ear image. Value mode feature.

Finally, the gradient direction histogram feature and the feature vector of the local binary pattern feature are cascaded to obtain a new feature vector to realize feature fusion, and input it to the minimum distance classifier for classification and identification.

The specific method is as follows: input the feature vector obtained by processing the above-mentioned steps of the human ear image to be recognized and several known classified human ear images respectively into the minimum distance classifier, and calculate the feature vector of the human ear image to be recognized and each of the recognized human ear images respectively. The distance between the feature vectors of the classified human ear images is known, and the classification corresponding to the smallest distance is the classification of the human ear image to be recognized. The distance calculation formula is as follows:

where d(i) is the distance between the human ear image to be recognized and the i-th known-classified human ear image, t ₁ (i) is the first feature vector of the i-th known-classified human ear image, p ₁ is the first feature vector of the human ear image to be recognized, t _n (i) is the n-th feature vector of the i-th known classified image, and p _n is the n-th feature vector of the image to be recognized.

The above is only the preferred embodiment of the present invention, it should be pointed out: for those skilled in the art, under the premise of not departing from the principle of the present invention, several improvements and modifications can also be made, and these improvements and modifications are also It should be regarded as the protection scope of the present invention.

Claims (3)

1. a human ear recognition method based on gradient direction histogram and local binary pattern fusion, is characterized in that, comprises the following steps: Step 1) collect several human ear images of known classification; Step 2) Calculate the feature value on each pixel in the human ear image, obtain the gradient direction histogram feature of the human ear image after dividing and standardizing; Step 3) Use the principal component analysis method to spatially transform the gradient direction histogram of the human ear image, so that the original coordinates are projected to a new space with lower dimensions and mutually orthogonal, so as to realize the gradient direction of the human ear image. Dimensionality reduction of histogram features; Step 4) calculating the local binary pattern value of each pixel on the human ear image to obtain the local binary pattern feature of the human ear image; Step 5) cascade the feature of the gradient direction histogram feature after the dimension reduction and the feature of the local binary pattern feature, obtain a new feature vector, and realize feature fusion; Step 6) Input the feature vector obtained after the processing of the human ear image to be recognized and several known classified human ear images into the minimum distance classifier, and calculate the feature vector of the human ear image to be recognized and The distance between the feature vectors of each known classified human ear image, the classification corresponding to the minimum distance is the classification of the human ear image to be recognized; Described step 2) is as follows: Step 21) Perform color standardization on the human ear image obtained in step 1), and uniformly convert the image into a grayscale image, and the conversion formula is: H(x, y)=0.3*R(x,y)+0.59* G(x, y)+0.11*B(x, y), where R(x, y), G(x, y), B(x, y) are the red, green, The color value of blue, H(x, y) represents the gray value of each pixel; Step 22) Use the Sobel operator to calculate the modulus value and direction angle of the gradient of each pixel of the image:

Among them, G(x, y) represents the gradient magnitude of the pixel point, α(x, y) represents the gradient direction of the pixel point, and H(x, y) represents the gray value of the pixel point; Step 23) Space and direction cells are voted with weights; first calculate the direction weights: x(i)=cos(θ), y(i)=sin(θ), θ=θ+π/(N _direction +1) , where: i is the direction number; θ is the angle, the initial value is 0; x(i) is the weight of the x-axis difference in the i direction; y(i) is the weight of the y-axis difference in the i direction, and N _direction is the total number of directions; Step 24) calculate the amplitude and direction, the amplitude is the mean square value of the image difference of the x or y axis, and the direction value is taken as the weighted maximum value in each direction; Step 25) Building blocks and standardizing, summarizing and combining the features in the cells into blocks, the calculation method is:

Among them, B(x) and B(y) represent the total value of the x-axis and the total value of the y-axis of the block respectively, C(x) and C(y) represent the total value of the x-axis and the total value of the y-axis of the cell, and B(size) is the size of the block, and B(step) is the step size of the block change; Step 26) Summarize the feature values in different directions and blocks, and construct the direction gradient histogram of the image. 2. a kind of human ear recognition method based on gradient direction histogram and local binary pattern fusion according to claim 1, is characterized in that, described step 3) is as follows: Step 31) Calculate the mean value of the gradient direction histogram feature of the corresponding pixel in the human ear image

Step 32) According to

Calculate the covariance matrix, where x _i is the feature that needs to be reduced in dimension,

is the mean value of the histogram feature of the gradient direction of the human ear image, and U ^T is the covariance matrix; Step 33) Take the first p principal components of the covariance matrix, and perform feature dimension reduction on the eigenvalues of each gradient direction histogram in the human ear image to obtain the gradient direction histogram feature of the dimension reduction by the principal component analysis method of the human ear image, the vector The dimension is p-dimension; the feature dimension reduction method is

where y represents the principal component feature. 3. a kind of human ear recognition method based on gradient direction histogram and local binary pattern fusion according to claim 1, is characterized in that, described step 4) is specifically as follows: Step 41) Define the central pixel of the 3×3 window as the threshold value, compare the gray values of the remaining 8 pixels with the threshold value in turn, and mark the pixel larger than the central pixel as 1, otherwise mark as 0, and form an 8-bit binary number. the binary representation of the local binary mode value of the window;

where (x _c , y _c ) represents the central element of the 3×3 neighborhood, its pixel value is _ic , _ip represents the value of other pixels in the neighborhood, s(x) is the sign function, and when x≥0, it is 1, otherwise 0, LBP(x _c , y _c ) is the binary representation of the local binary pattern value of the center pixel; Step 42) Convert the LBP (x _c , y _c ) of each pixel into decimal numbers to obtain the final local binary pattern value, and after summarizing the local binary pattern features of the human ear image.

Images