Sika deer individual identification method based on computer vision

Abstract

The invention discloses a sika deer individual identification method based on computer vision, and relates to the field of computer vision, in particular to a sika deer individual identification method, which can design a sika deer individual identification model by utilizing a method of combining a convolutional neural network and Transform. More detailed spot feature information is extracted from sika deer data set images, so that the capability of identifying sika deer individuals in a real sika deer breeding environment can be improved. According to the technical scheme, firstly, sika deer spots serve as features, and a sika deer individual data set is established; secondly, designing a sika deer individual identification model based on computer vision by adopting a method of combining a convolutional neural network and a Transform, and realizing non-contact identification of the sika deer; thirdly, training and verifying the sika deer individual data set; experimental results show that the model can accurately, efficiently and intelligently identify sika deer individuals, and is suitable for solving the key problem of sika deer individual identification in an actual deer field. The method can be applied to individual identification of the sika deer in an actual sika deer farm, and intelligent and informationized breeding of the sika deer in the sika deer farm can be promoted.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision, and in particular to a computer vision-based sika deer individual identification method. BACKGROUND

[0002] Sika deer has very high economic value, medicinal value and social value. With the increasing demand for deer products in the market, the development of sika deer breeding industry is promoted. At the same time, the health management of sika deer farms is gradually moving towards informatization, precision and intelligence. Sika deer has been listed in the "National Livestock and Poultry Genetic Resources Catalogue", therefore, it is necessary to strengthen the breeding of good varieties, health management, scientific feeding, timely breeding and selection of superior breeding. Sika deer individual identification is the basis for sika deer seed source breeding, therefore, sika deer individual identification is of great significance for sika deer breeding, management and seed source breeding. In recent years, non-contact animal individual identification technology as a new research direction in the field of computer vision has the advantages of low cost, high reliability and the like. Compared with ear tags and ear numbers, non-contact sika deer individual identification technology is more beneficial to the health of sika deer, therefore, it is necessary to propose an efficient, non-contact and practical sika deer individual identification method for sika deer farms to realize precise sika deer breeding. SUMMARY

[0003] In view of the deficiencies in the prior art, the present application proposes a computer vision-based sika deer individual identification method, which can use a method combining convolutional neural networks and Transformers to design a sika deer individual identification model, extract more detailed spot feature information from sika deer dataset images, and improve the identification ability of sika deer individuals in a real sika deer breeding environment. The application adopts the following technical means: in order to solve the above technical problems or at least partially solve the above technical problems, the present application provides a sika deer individual identification method based on the combination of convolutional neural networks and Transformers.

[0004] The present application provides a sika deer individual identification method based on the combination of convolutional neural networks and Transformers, characterized in that it comprises the following steps: Step (1), sika deer image acquisition; collect individual image data corresponding to multiple sika deer, collect sika deer side spots as features, and save to a sika deer individual dataset according to sika deer individual labels respectively; Step (2), image segmentation; image segmentation is performed on each individual image of the target sika deer in the sika deer individual dataset having side spots, to obtain a plurality of single-background sika deer individual pictures having sika deer side spots; Step (3), based on the multiple single background individual pictures of sika deer, a sika deer individual training dataset for model training is constructed, the sika deer individual training dataset is divided into a training dataset and a validation dataset, and a test dataset is established at the same time; Step (4), image preprocessing; data augmentation is performed on each sika deer individual picture in the sika deer individual training dataset to obtain a data-augmented training dataset, a validation dataset and a test dataset; Step (5), constructing a sika deer individual recognition network model; a sika deer individual recognition network model for extracting sika deer side-view spot features is constructed by combining convolutional neural network and Transformer, and AM-Softmax is used as the loss function of the output layer of the sika deer individual recognition network model; Step (6), the sika deer individual network model is trained using the data-augmented sika deer individual training dataset and the parameters are optimized, and the network weight parameters are obtained through the validation of the data-augmented validation dataset; Step (7), during testing, the sika deer individual recognition network model is tested using the data-augmented test dataset, the model parameters of the sika deer individual recognition network model are the network weight parameters, and the sika deer individual recognition network model is determined to be trained when the testing is passed; Step (8), during recognition, the sika deer individual recognition network model is input with the collected sika deer individual picture with side-view spots, and the sika deer individual recognition network model outputs the recognition result, which includes the sika deer individual label.

[0005] The specific implementation method of step (1) is to collect individual image data of sika deer, which can be achieved by using ordinary cameras or mobile phones and other devices to take pictures of sika deer in sika deer farms to obtain original image data of sika deer. In order to ensure the quality and diversity of image data, pictures taken under different angles, distances, lighting conditions, etc. can be stored in folders corresponding to the number of sika deer individual labels of the target sika deer for recognition.

[0006] The specific implementation method of step (2) is to use DeepLabV3 image segmentation model to perform image segmentation on the individual images with side-view spots in the sika deer individual dataset to obtain a sika deer individual dataset with single background and interference removed.

[0007] The specific implementation method of the step (3) is that the individual data set of the sika deer obtained in the step (2) is randomly divided into a training data set and a verification data set according to a ratio of 8:2, and part of the clear image of each sika deer is selected from the entire sika deer individual data set as a test data set, and the sika deer individual test data set is used for testing.

[0008] The specific implementation method of the step (4) is that the data set obtained in the step (3) is subjected to data enhancement on the sika deer individual data set after removing the background through horizontal flipping, vertical flipping, brightness enhancement, graying and the like, so as to obtain the training data set, the verification data set and the test data set after data enhancement; the data enhancement can effectively reduce the misrecognition due to the positional deviation, thereby effectively increasing the recognition accuracy of the sika deer individual.

[0009] The specific implementation method of the step (5) is that the network combining the convolutional neural network and the Transformer is designed as a feature extraction module in the style of the Transformer, including four ConvBlock modules and three IDBlock modules, and all the ConvBlock modules and the IDBlock modules contain MS-CoTBlock modules with the same structure; the input image is first input into a 64-channel, 7*7-sized cavity convolutional layer which is formed by a 3*3-pixel-sized ordinary convolutional layer with a 3-fold expansion and has a large convolutional kernel receptive field; then, the image is input into a 3*3-sized maximum pooling layer with a step of 2 for feature extraction, so as to compress the image, and then the first ConvBlock, the first IDBlock, the second ConvBlock, the second IDBlock, the third ConvBlock, the third IDBlock and the fourth ConvBlock are sequentially connected, and an SE module is connected after each IDBlock to compress and excite the obtained feature information to extract the target features of the sika deer spots and perform feature extraction on the target features; then, a global average pooling layer with a step of 1 and a size of 7*7 and a channel number of 2048 is used to optimize the network structure, the multi-dimensional feature matrix is changed into a one-dimensional feature sequence through a Flatten layer, and is input into a fully connected layer, in order to prevent overfitting, a 3*3-sized convolutional layer C1 with a channel number of 2048 is connected in the fully connected layer, and then a maximum layer P1 with a pixel size of 3*3 and a step of 1 is used, and then two fully connected layers F1 and F2 are connected, with 2048 and 512 neurons respectively connected. The generalization and anti-overfitting ability of the model are increased; a 2048-dimensional feature vector is obtained through the fully connected layer; finally, a Softmax classifier is used to output the sika deer individual category; The number of modules in each layer is 2, 2, 3 and 1 respectively, the MS-CoTBlock module is used for feature extraction of important features, and the Softmax classification layer is used for classification.

[0010] The short connection part in the ConvBlock uses a max pooling layer with a 3*3 kernel and a step of 2, which is connected with a convolution layer with a 1*1 kernel and a step of 1 and a batch normalization layer, so as to reduce the information loss during network training, and the sika deer individual identification network model uses a GELU activation function.

[0011] The specific implementation method of the step (6) is that the sika deer individual identification network model is trained using the training data set, the convergence of the loss function value of the sika deer individual identification network model during training is monitored, if the loss function value is greater than a set threshold, the network model parameters are adjusted, and if the loss function value is less than the set threshold and stable in the set interval, the network weight parameters are obtained.

[0012] The specific implementation method of the step (7) is that the sika deer individual picture with a side-view spot in the test data set is input into the sika deer individual identification network model, the sika deer individual identification network model outputs a prediction result, the prediction result is compared with the sika deer individual label corresponding to the sika deer individual picture with a side-view spot in the test data set, if the prediction result matches the sika deer individual label in the test data set, it is determined that the test is passed, and the training is completed; if the identification is wrong, the step (6) is returned to continue training until a trained network model is obtained.

[0013] The specific implementation method of the step (8) is that the sika deer individual picture with a side-view spot collected is input into the sika deer individual identification network model, the sika deer individual identification network model outputs an identification result, and the identification result includes a sika deer individual label.

[0014] The above technical solutions provided by the embodiments of the present application have the following advantages compared with the prior art: The embodiments of the present application use the DeepLabV3 image segmentation model to perform image segmentation on each sika deer image with a side-view spot in the sika deer individual data set, so as to obtain a sika deer individual segmentation data set without any interference of a single background. The sika deer individual identification model based on the combination of a convolutional neural network and a Transformer is trained using the sika deer individual data set without background interference, the sika deer individual is identified without contact, the sika deer is not harmed, the stress of the sika deer is effectively avoided, the identification efficiency is high, and the cost is low.

[0015] The present application has the following advantages: The application designs reasonably, adopts a mode of combining convolutional neural network and Transformer to construct a recognition model, solves the problem that there is no available spatial information in the convolutional neural network, and after different textures are integrated in the image, the recognition of the network will deviate, introduces a compression module to reduce the size of the original image, solves the time-consuming problem in model training, and retains the image features. Under the real breeding environment of sika deer, the network constructed by combining convolutional neural network and Transformer can extract more detailed spot information from the sika deer dataset image, enhance the recognition ability, and has reference significance for recognizing animal individuals with high similarity.

[0016] The drawings incorporated into the specification and forming a part thereof show embodiments consistent with the present application and, together with the specification, serve to explain the principles of the application.

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows, and obviously, other drawings can also be obtained by those skilled in the art without creative labor under the premise of not paying creative labor.

[0018] Figure 1 The overall implementation flowchart of the present application is shown in the figure. Figure 2 The structure schematic diagram of the present application containing MS-CoTNet network structure is shown in the figure. Figure 3 The structure schematic diagram of the present application containing ConvBlock structure is shown in the figure. Figure 4 The structure schematic diagram of the present application containing IDBlock structure is shown in the figure.

[0019] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the present application.

[0020] At present, computer vision-based non-contact animal individual identification technology as a new research direction in the field of artificial intelligence has become a hot spot of scholars. Contrary to direct visual observation and manual monitoring, this method saves labor costs and is feasible in large-scale livestock farming. In addition, compared with ear tags, ear numbers and electronic sensors, the computer vision-based animal individual identification method does not have the disadvantages of measurement noise due to sensor failure, sensor damage or loss due to non-contact with animals, and harm to animals; on the contrary, this method not only effectively avoids the stress response of animals, but also is more beneficial to animal health, so it has attracted people's attention. Through individual identification technology to identify the individual of the farm animals, in order to achieve the purpose of tracking breeding, therefore, it is of great significance to use animal individual identification technology to monitor it without contact. Therefore, the embodiment of the present application provides a sika deer individual identification method combining convolutional neural network and Transformer.

[0021] Figure 1 The flowchart of the whole implementation of the present application comprises: S110: sika deer image acquisition; S120: image segmentation; S130: sika deer individual dataset construction; S140: image preprocessing; S150: sika deer individual identification network model construction; S160: individual identification result; The embodiment of the present application provides a sika deer individual identification method combining convolutional neural network and Transformer; the method comprises: Step (1), sika deer image acquisition.

[0022] The individual image data of sika deer can be collected by using ordinary cameras or mobile phones and other devices to take pictures of sika deer in the sika deer farm, and the original image data of sika deer is obtained. In order to ensure the quality and diversity of the image data, pictures taken under different angles, distances, lighting conditions, etc. can be stored in a folder corresponding to the number of the target sika deer individual label for identification.

[0023] Step (2), image segmentation.

[0024] The individual image of each target sika deer in the sika deer individual dataset with side-view spots is segmented to obtain a plurality of sika deer individual pictures with side-view spots of sika deer in a single background; the individual image with side-view spots in the sika deer individual dataset is segmented by using a DeepLabV3 image segmentation model to obtain a sika deer individual dataset with a single background and interference removed.

[0025] Step (3), constructing the individual data set of sika deer, based on a plurality of single background sika deer individual pictures, constructing the sika deer individual training data set for model training, dividing the sika deer individual data set obtained in step (2) into training data set and verification data set according to the ratio of 8:2, and selecting part of clear images of each sika deer from the whole sika deer individual data set as test data set, which is used for testing.

[0026] Step (4), image preprocessing.

[0027] The sika deer individual data set after removing the background is data enhanced through horizontal flip, vertical flip, brightness enhancement, graying and the like, so as to obtain the training data set, the verification data set and the test data set after data enhancement; the data enhancement can effectively reduce the misrecognition caused by position deviation, thereby effectively increasing the recognition accuracy of the sika deer individual.

[0028] Step (5), constructing the sika deer individual recognition network model.

[0029] Figure 2 The structure diagram of the MS-CoTNet network structure contained in the application, the method is based on the combination of convolutional neural network and Transformer to construct the sika deer individual recognition network model for extracting sika deer spot features, and AM-Softmax is used as the loss function of the output layer of the sika deer individual recognition network model; The network combining convolutional neural network and Transformer is a feature extraction module designed in the style of Transformer, including four ConvBlock modules and three IDBlock modules, and all ConvBlock modules and IDBlock modules contain MS-CoTBlock modules with the same structure; the input image is first input into a hole convolution layer with 64 channels and a size of 7×7, which is formed by expanding a normal convolution of 3×3 pixels by 3 times and has a receptive field of a large convolution kernel; then it enters a 3×3 maximum pooling layer with a stride of 2 to extract features from the image to compress the image, followed by the first ConvBlock, the first IDBlock, the second ConvBlock, the second IDBlock, and the like. lock, the third ConvBlock, the third IDBlock, and the fourth ConvBlock. After each IDBlock, the SE module is connected to compress and excite the obtained feature information to extract the target features of the sika deer spots, and then extract the target features; then, a global average pooling layer with a stride of 1, a size of 7×7, and a number of channels of 2048 is used to optimize the network structure. The multi-dimensional feature matrix is ​​converted into a one-dimensional feature sequence through the Flatten layer and imported into the fully connected layer. To prevent overfitting, a 3×3 convolutional layer C1 with a channel number of 2048 is connected to the fully connected layer. Then, the pixel size of the maximum layer P1 is 3×3 and the stride is 1. Then, two fully connected layers F1 and F2 are connected, with 2048 and 512 neurons respectively. This increases the generalization and anti-overfitting capabilities of the model; a 2048 feature vector is obtained through the fully connected layer; finally, a Softmax classifier is used to output the individual category of the sika deer. The number of modules in each layer is 2, 2, 3, and 1 respectively. The MS-CoTBlock module is used for feature extraction of important features, and the Softmax classification layer is used for classification.

[0030] The ConvBlock module of the present invention is as follows Figure 3 As shown in the figure, it includes convolution kernel (Conv2d), normalization layer (BatchNorm2d), activation layer (GELU), maximum pooling layer (Maxpool) and multi-scale feature extraction layer (MS-CoTBlock). The main path of each ConvBlock module includes convolution kernel and multi-scale feature extraction layer. The shortcut path uses the maximum pooling layer in combination with 1×1 convolution to help improve the translation invariance of the network. The IDBlock module is as follows Figure 4As shown, the IDBlock module is followed by an SE module, which contains a convolution kernel (Conv2d), a normalization layer (BatchNorm2d), an activation layer (GELU), and a multi-scale feature extraction layer (MS-CoTBlock).

[0031] Step (6), the training data set is used to train the sika deer individual identification network model, and the convergence of the loss function value of the sika deer individual identification network model during training is monitored. If the loss function value is greater than the set threshold, the network model parameters are adjusted. If the loss function value is less than the set threshold and stable in the set interval, the network weight parameters are obtained.

[0032] Step (7), during the test, the sika deer individual picture with side-view spots in the test data set is input into the sika deer individual identification network model, and the sika deer individual identification network model outputs a prediction result. The prediction result is compared with the sika deer individual label corresponding to the sika deer individual picture with side-view spots in the test data set. If the prediction result matches the sika deer individual label in the test data set, it is determined that the test is passed, and the training is completed. If the identification is wrong, return to step (6) for continuous training until a trained network model is obtained.

[0033] Step (8), during identification, the collected sika deer individual picture with side-view spots is input into the sika deer individual identification network model, and the sika deer individual identification network model outputs an identification result, which includes a sika deer individual label.

[0034] The above examples are combined Figure 1 is a specific use flowchart of the present application; Figure 2 is a structural schematic diagram of the MS-CoTNet network structure of the present application; Figure 3 is a structural schematic diagram of the ConvBlock structure of the present application; Figure 4 is a structural schematic diagram of the IDBlock structure of the present application. The present application realizes the classification of sika deer individuals by constructing a sika deer individual identification network model for extracting sika deer spot features in the manner of combining a convolutional neural network with a convolutional neural network and a Transformer.

[0035] It has to be noted that, in the present document, relational terms are intended only to convey a possible relationship between elements or

[0036] The above description is merely that of the specific embodiments of the application and as such is not to be taken in a limiting sense. Various modifications and co nti n uations will be evident to those skilled in the art that do not depart from the spirit and scope of the application as defined by the appended claims. The scope of the application should be determined by the scope of the appended claims and their equivalents.

Claims

1. A computer vision-based sika deer individual identification method, characterized by; The method comprises the following steps: Step (1), image collection of sika deer; Collect individual image data of multi-antlered sika deer, collect sika deer side spot features, and save individual data sets of sika deer according to individual labels of sika deer; Step (2), image segmentation; Image segmentation is performed on each individual image of the target sika deer in the sika deer individual data set with side spots to obtain a plurality of single-background sika deer individual pictures with sika deer side spots; Step (3), based on a plurality of single-background sika deer individual pictures, a sika deer individual training data set for model training is constructed, the sika deer individual training data set is divided into a training data set and a verification data set, and a test data set is established at the same time; Step (4), image preprocessing; Data augmentation is performed on each sika deer individual picture in the sika deer individual training data set to obtain a training data set, a verification data set and a test data set after data augmentation; Step (5), constructing a sika deer individual recognition network model; A sika deer individual recognition network model for extracting sika deer side spot features is constructed by combining convolutional neural network and Transformer, and AM-Softmax is used as the loss function of the output layer of the sika deer individual recognition network model; Step (6), the sika deer individual network model is trained using the sika deer individual training data set after data augmentation and the parameters are optimized, and the network weight parameters are obtained through the verification data set after data augmentation; Step (7), during testing, the sika deer individual recognition network model is tested using the test data set after data augmentation, the model parameters of the sika deer individual recognition network model are the network weight parameters, and the sika deer individual recognition network model is determined to be trained when the test is passed; Step (8), during recognition, the sika deer individual picture with side spots collected is input into the sika deer individual recognition network model, and the sika deer individual recognition network model outputs the recognition result, wherein the recognition result includes the sika deer individual label.

2. The method of claim 1, wherein the method is characterized by, The specific implementation method of step (1) is: Collect individual image data of sika deer, which can be obtained by using ordinary cameras or mobile phones and other devices to take pictures of sika deer in sika deer farms, and obtaining original image data of sika deer. In order to ensure the quality and diversity of image data, pictures taken under different angles, distances, lighting conditions, etc. can be stored in folders corresponding to the sika deer individual label of the target sika deer for identification.

3. The specific implementation method of step (2) is to use DeepLabV3 image segmentation model to perform image segmentation on the individual image with side spots in the sika deer individual data set to obtain a single-background sika deer individual data set without interference.

4. The specific implementation method of step (3) is to randomly divide the sika deer individual data set obtained in step (2) into a training data set and a validation data set according to a ratio of 8:2, and select part of the clear image of each sika deer from the entire sika deer individual data set as a test data set, which is used for testing.

5. The specific implementation method of step (4) is to perform data enhancement on the sika deer individual data set after removing the background by horizontal flipping, vertical flipping, brightness enhancement, and grayscale method, and obtain the data enhanced training data set, validation data set and test data set; data enhancement can effectively reduce the misrecognition caused by position deviation, thereby effectively increasing the recognition accuracy of sika deer individuals.

6. The specific implementation method of step (5) is that the network combining the convolutional neural network and the Transformer is designed as a feature extraction module in the style of Transformer, including four ConvBlock modules and three IDBlock modules, and all ConvBlock modules and IDBlock modules contain MS-CoTBlock modules with the same structure; the input image is first input into a 64-channel, 7×7-sized empty convolutional layer, which is formed by a 3×3-pixel ordinary convolutional layer with a 3-fold expansion, and has a large receptive field of convolution kernel; then it enters a 3×3 maximum pooling layer with a step of 2, which extracts features from the picture to compress the picture, followed by a first ConvBlock, a first IDBlock, a second ConvBlock, a second IDBlock, a third ConvBlock, a third IDBlock, and a fourth ConvBlock, and an SE module is connected after each IDBlock to compress and excite the obtained feature information to extract the target features of sika deer spots and extract the target features; then a global average pooling layer with a step of 1, a size of 7×7, and a channel number of 2048 is used to optimize the network structure, and the multi-dimensional feature matrix is converted into a one-dimensional feature number sequence through the Flatten layer, and is imported into the fully connected layer. In order to prevent overfitting, a 3×3 convolutional layer C1 with a channel number of 2048 is connected in the fully connected layer, and then a maximum layer P1 with a pixel size of 3×3 and a step of 1 is connected, followed by two fully connected layers F1 and F2 with 2048 and 512 neurons respectively. Increase the generalization and anti-overfitting ability of the model; a 2048 feature vector is obtained through the fully connected layer; finally, a Softmax classifier is used to output the sika deer individual category; The number of modules in each layer is 2, 2, 3 and 1 respectively, the MS-CoTBlock module is used for feature extraction of important features, and the Softmax classification layer is used for classification.

7. The specific implementation method of step (5) is that the short connection part in the ConvBlock uses a max pooling layer with a 3*3 kernel and a step of 2, the max pooling layer is connected with a convolution layer with a 1*1 kernel and a step of 1 and a batch normalization layer, the information loss during network training is reduced, and the sika deer individual identification network model uses a GELU activation function. The specific implementation method of step (6) is that the sika deer individual identification network model is trained using the training data set, the convergence condition of the loss function value of the sika deer individual identification network model during training is monitored, if the loss function value is greater than a set threshold, the network model parameters are adjusted, if the loss function value is less than the set threshold and stable in the set interval, the network weight parameters are obtained.

8. The specific implementation method of step (7) is that the sika deer individual picture with side-view spots in the test data set is input into the sika deer individual identification network model, the sika deer individual identification network model outputs a prediction result, the prediction result is compared with the sika deer individual label corresponding to the sika deer individual picture with side-view spots in the test data set, if the prediction result matches the sika deer individual label in the test data set, it is determined that the test is passed, and the training is completed; if the identification is wrong, step (6) is returned to continue training until a trained network model is obtained.

9. The specific implementation method of step (8) is that the collected sika deer individual picture with side-view spots is input into the sika deer individual identification network model, the sika deer individual identification network model outputs an identification result, and the identification result includes a sika deer individual label.

Images