CN108197605A - Yak personal identification method based on deep learning - Google Patents

Abstract

The invention discloses a yak identity recognition method based on deep learning, comprising the following steps: S1, collecting yak pictures and videos, and decoding the videos into pictures; S2, performing classification regression and position on the pictures through the target detection network Faster R-CNN Regression, obtain the pixel position of the cow face on the picture and its confidence degree; Then the obtained cow face pixel position is cut out, and the cow face is intercepted; S3, the cow face obtained in step S2 is input into the feature extraction network, and the cow face is processed Feature extraction, and output the corresponding feature vector; S4, match the feature vector extracted in step S3 with the cow face feature vector in the database, calculate the similarity of the cow face, and output the one with the highest similarity with the yak in the database A picture of a yak to complete the yak identification. The invention avoids the uncertainty of manual feature extraction and the limitation of the ear tag identification method in the traditional identification method, effectively improves the efficiency of yak identification, and can reduce the risk of insurance fraud.

Description

Yak identification method based on deep learning

technical field

The invention belongs to the field of deep learning image processing and the field of yak identification, and in particular relates to a yak identification method based on deep learning.

Background technique

With the advancement of science and technology and the development of society, the situation of livestock insurance is becoming more and more common. Yaks, as the main source of income for herdsmen in the plateau area, are widely raised in this area. However, the ability of herdsmen to resist natural disasters and epidemic risks is low, and the yak breeding industry is facing huge risks. Therefore, yak insurance is also actively promoted in the yak breeding industry, which brings the risk of insurance fraud. Currently, insurance fraud is spreading in China.

Traditional yak identification is achieved by earmarking the yak. This method has many disadvantages. First, the ear tags are highly reproducible and there will be no biological differences. Second, the ear tags can be reused, which cannot prevent the reuse of ear tags after death of yaks. Therefore, a more accurate and effective yak identification method is needed.

Contents of the invention

The purpose of the present invention is to overcome the uncertainty of manual feature extraction and the limitations of ear tag identification methods in the existing identification methods, and provide a yak based on deep learning that can effectively improve the efficiency of yak identification and reduce the risk of insurance fraud. Identification method.

The object of the present invention is achieved through the following technical solutions: the yak identification method based on deep learning comprises the following steps:

S1, collecting yak pictures and videos, and decoding the videos into pictures;

S2. Perform classification regression and position regression on the collected yak picture through the target detection network Faster R-CNN to obtain the pixel position and confidence of the cow face on the picture; then crop the obtained cow face pixel position and intercept it cow face;

S3. Input the cow face obtained in step S2 into the feature extraction network, perform feature extraction on the cow face, and output the corresponding feature vector;

S4. Match the feature vector extracted in step S3 with the cow face feature vector in the database, calculate the similarity of the cow face, output a yak picture in the database with the highest similarity with the yak, and complete the yak recognition.

Further, the step S2 includes the following sub-steps:

S21. Normalize the image to a size of 224*224;

S22. Through 13 convolutional layers and 5 times of downsampling, 512 feature maps with a size of 14*14 are obtained;

S23. Perform the following processing on each feature map: slide on the feature map through a convolution kernel of 3*3 size, use the center of each convolution kernel as a reference point, and then select 3 different area sizes and sizes around the reference point. 3 different size ratios, generating 9 candidate regions;

S24, remove the candidate frame mapped to the original image beyond the boundary of the original image in the candidate area;

S25. Map the candidate area to the feature map of the last layer of network convolution, generate a fixed-size feature map for each candidate area through the ROI pooling layer, classify and refine the position of the generated feature map, and generate in combination with step S1 The size of the picture is calculated to obtain the pixel position of the cow face on the picture and its confidence level;

S26. Cut out the obtained cow face pixel position, and intercept the cow face.

Further, the step S3 includes the following sub-steps:

S31. Adjust the cow face picture to a size of 224*224, and then input the feature extraction network for feature extraction;

S32. The feature extraction network outputs a 4096-dimensional feature vector through a convolutional layer, a pooling layer, and a fully connected layer.

Further, the step S4 includes the following sub-steps:

S41, compare the feature vector extracted in step S3 with the data in the cow face library, and calculate the similarity, wherein the similarity calculation adopts the method of cos angle:

Among them, cosθ represents the cosine of the angle between the two feature vectors, and a and b respectively represent the feature vector extracted by S2 and the feature vector in the cow face database;

S42. Output the cosθ angle value of the two yaks as the similarity, and output the serial number corresponding to the yak with the greatest similarity with the yak to be identified in the database as the yak identification result.

The beneficial effects of the present invention are: the present invention detects the cow face of yaks through the method based on deep learning, then automatically extracts the features of the cow face with a convolutional neural network, and finally calculates the similarity between two cow faces by using the cos angle, and outputs The yak number with the highest similarity is used as the matching result, which can achieve a relatively high accuracy rate. It avoids the uncertainty of manual feature extraction in traditional identification methods and the limitations of ear tag identification methods, effectively improves the efficiency of yak identification, and can reduce the risk of insurance fraud.

Description of drawings

Fig. 1 is the flowchart of the yak identification method based on deep learning of the present invention;

Fig. 2 is the cow face picture collected in the present embodiment;

Fig. 3 is the bull's face picture that cuts out in the present embodiment;

FIG. 4 is a picture of the output result of cow face recognition in this embodiment.

Detailed ways

The technical solution of the present invention will be further described below in conjunction with the accompanying drawings.

As shown in Figure 1, the yak identification method based on deep learning includes the following steps:

S1, collecting yak pictures and videos, and decoding the videos into pictures;

S2. Perform classification regression and position regression on the collected yak picture through the target detection network Faster R-CNN to obtain the pixel position and confidence of the cow face on the picture; then crop the obtained cow face pixel position and intercept it Cow face; specifically includes the following sub-steps:

S21. Normalize the image to a size of 224*224;

S22. Through 13 convolutional layers and 5 times of downsampling, 512 feature maps with a size of 14*14 are obtained;

S23. Perform the following processing on each feature map: slide a 3*3 convolution kernel on the feature map, set an anchor mechanism, that is, use the center of each convolution kernel as a reference point, Then select 3 different area sizes around the reference point (128, 256, 512, corresponding to feature maps are 3, 6, 12) and 3 different size ratios (1:1, 1:2 and 2:1), generate 9 candidate areas;

S24, remove the candidate frame mapped to the original image beyond the boundary of the original image in the candidate area;

S25. Map the candidate area to the feature map of the last layer of network convolution, generate a fixed-size feature map for each candidate area through the ROI pooling layer, classify and refine the position of the generated feature map, and generate in combination with step S1 The size of the picture is calculated to obtain the pixel position of the cow face on the picture and its confidence level;

S26. Cut out the obtained cow face pixel position, and intercept the cow face.

S3. Input the cow face obtained in step S2 into the feature extraction network, perform feature extraction on the cow face, and output the corresponding feature vector; including the following substeps:

S31. Adjust the cow face picture to a size of 224*224, and then input the feature extraction network for feature extraction;

S32. The feature extraction network outputs a 4096-dimensional feature vector through a convolutional layer, a pooling layer, and a fully connected layer.

S4, match the feature vector extracted in step S3 with the cow face feature vector in the database, calculate the similarity of the cow face, output a yak picture with the highest similarity with the yak in the database, and complete the yak recognition; specifically include The following substeps:

S41, compare the feature vector extracted in step S2 with the data in the cow face database, and calculate the similarity, wherein the similarity calculation adopts the method of cos angle:

Among them, cosθ represents the cosine of the angle between the two feature vectors, and a and b respectively represent the feature vector extracted by S2 and the feature vector in the cow face database;

S42. Output the cosθ angle value of the two yaks as the similarity, and output the serial number corresponding to the yak with the greatest similarity with the yak to be identified in the database as the yak identification result.

Applications:

1. Sample making: A total of 3988 yak pictures were collected in this embodiment. 2900 images are randomly selected as model training samples. Make a target detection sample and use the annotation tool to mark the positions of all cow faces in the picture.

2. Training stage:

(1) Design the target detection model Faster R-CNN based on the deep convolutional network VGG16-Net, and use the selected 2900 pictures as training samples to train the model, with a total of 60,000 iterations; the target detection model Faster R-CNN is commonly used in this field A target detection model, the specific training process of which will not be repeated here.

(2) Design the deep convolutional network VGG16-Net, and use the cropped cow face training model for a total of 100,000 iterations.

3. Testing phase:

(1) Make a yak database containing 315 yaks, and then randomly select 455 yak samples for accuracy and similarity testing;

(2) Input 455 yak samples into the trained Faster R-CNN network model, and output the detected cow face pictures, as shown in Figure 2. According to the coordinates of the cow face output by the target detection, the corresponding cow face is cut out, as shown in Figure 3;

(3) input the cow face feature extraction model with the cropped cow face, and output a 4096-dimensional feature vector;

(4) Perform similarity detection according to the eigenvector output in (3) and the yak data in the database, and finally output the number of the yak as the recognition result, as shown in Figure 4, which are the similarity between the test picture and the database The highest (0.923420) cow face image. There were 416 test samples in total, of which 31 were erroneously recognized yak pictures, and the recognition accuracy rate was 0.92548.

In summary, it can be seen that the present invention firstly detects the cow face of the yak through the method based on deep learning, then uses the convolutional neural network to automatically extract the cow face features, and finally uses the cos angle to calculate the similarity of the two cow faces, and outputs The yak number with the highest similarity is used as the matching result, which can achieve a relatively high accuracy rate. It avoids the uncertainty of manually extracting features in the traditional identification method, and the limitation of using ear tags for yak identification, and effectively improves the efficiency of yak identification.

Those skilled in the art will appreciate that the embodiments described here are to help readers understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical revelations disclosed in the present invention without departing from the essence of the present invention, and these modifications and combinations are still within the protection scope of the present invention.

Claims (4)

1. the yak personal identification method based on deep learning, which is characterized in that include the following steps：

S1, acquisition yak picture and video, picture is decoded as by video；

S2, collected yak picture is subjected to classification recurrence and position recurrence by target detection network Faster R-CNN, Obtain location of pixels of the ox face on picture and its confidence level；Then obtained ox face location of pixels is cut, intercepted out Ox face；

S3, the ox face input feature vector for obtaining step S2 extract network, and feature extraction is carried out, and export corresponding feature to ox face Vector；

Ox face feature vector in S4, feature vector and database that step S3 is extracted matches, and calculates the phase of ox face Like degree, output database neutralizes the highest yak picture of similarity of the yak, completes yak identification.

2. the yak personal identification method according to claim 1 based on deep learning, which is characterized in that the step S2 Including following sub-step：

S21, picture is normalized to 224*224 sizes；

S22, by 13 convolutional layers, 5 down-samplings obtain the characteristic pattern of 512 size 14*14；

S23, each characteristic pattern is handled as follows：It is slided on characteristic pattern by the convolution kernel of 3*3 sizes, respectively with every Then 3 different area sizes and 3 kinds of different size ratios are chosen around datum mark in one convolution kernel center as a datum mark Example generates 9 candidate regions；

S24, remove be mapped in candidate region in artwork be more than artwork boundary candidate frame；

S25, candidate region is mapped on the characteristic pattern of last layer of convolution of network, each time is made for pooling layers by ROI Favored area generates fixed-size characteristic pattern, is classified to the characteristic pattern of generation and position refine, is generated with reference to step S1 Picture size is calculated, and obtains location of pixels and its confidence level that ox face is present on picture；

S26, obtained ox face location of pixels is cut, intercepts out ox face.

3. the yak personal identification method according to claim 2 based on deep learning, which is characterized in that the step S3 Including following sub-step：

S31, ox face picture is adjusted to 224*224 sizes, then input feature vector extraction network carries out feature extraction；

S32, feature extraction network export the feature vector of one 4096 dimension by convolutional layer, pond layer and full articulamentum.

4. the yak personal identification method according to claim 3 based on deep learning, which is characterized in that the step S4 Including following sub-step：

S41, the data in the step S3 feature vectors extracted and ox face library are compared, calculate similarity, wherein similarity It calculates by the way of cos angles：

Wherein, cos θ represent the folder cosine of an angle of two feature vectors, and a, b represent the feature vector and ox face data of S2 extractions respectively Feature vector in library；

S42, the cos θ angles values of two yaks are exported as similarity, and output in the database with yak similarity to be identified The corresponding number of maximum yak is as yak identification result.