Improved YOLOv8n-based behavior identification method for sheep raised in house

Abstract

The invention discloses an improved YOLOv8n-based behavior recognition method for sheep raised in a house, and relates to the field of sheep behavior recognition based on agricultural computer vision, and the method comprises the following steps: S1, obtaining behavior video data of sheep raised in a house; s2, preprocessing the video data, and constructing an image database of the sheep raised in a house; s3, on the basis of the images of the sheep raised in the house, sheep behaviors are marked through a marking tool; s4, dividing the processed image into a training set, a verification set and a test set, and expanding the data diversity of the training set by using a data augmentation mode; step S5, using a depth separable convolution to replace all standard convolution in the CSPDarkNet53 feature extraction network, and embedding a proposed PCBAM attention module into a neck network of YOLOv8n; s6, inputting the expanded training set into the improved YOLOv8n network model for training; and performing iteration by using a CioU loss function to obtain an optimal model.

Description

Technical Field

[0001] The present invention relates to the field of sheep behavior recognition in agricultural computer vision, and in particular to a method for identifying behaviors of housed sheep based on improved YOLOv8n. Background Art

[0002] Sheep farming has a profound history in China. Since ancient times, Chinese people have had the tradition of raising sheep for wool, mutton and other by-products. There are many sheep breeds in China, and there are dozens of well-known sheep breeds. While providing abundant living resources, sheep farming is also regarded as an important part of the agricultural ecosystem because of its adaptability to the environment, economic value and important role in ecological balance. The multiple benefits of sheep farming, such as providing high-quality meat, high-grade wool, biofertilizers and helping to protect grassland ecology, have made it known as a "multifunctional animal husbandry model for sustainable development" in modern society.

[0003] The behavior of sheep is often a direct reflection of their health status. For example, sheep with abnormal behavior often show loss of appetite, reduced activity, social isolation, etc. These behavioral characteristics may indicate the occurrence of disease and reproductive disorders. On the contrary, positive social interactions, regular eating and resting behaviors are usually accompanied by good health, stable weight gain, and high-quality wool production. However, traditional sheep behavior assessment mainly relies on manual observation and subjective judgment. This method is not only time-consuming and labor-intensive, but also easily affected by the observer's experience, emotions and environmental factors, resulting in insufficient reliability and consistency of the assessment results.

[0004] With the rapid development of deep learning technology, target detection algorithms based on deep learning have achieved certain results in some fields. For example, the Journal of the Chinese Society of Agricultural Machinery | Issue 4, 2024 | 231-240 | discloses a method for identifying the behavior of housed sheep based on improved YOLO v5n. This method proposes a sheep behavior recognition method, which is mainly used to solve the problem that the model has many parameters and is not conducive to deployment on edge devices. This patent document "Sheep face image detection method based on YOLOv8-Swin-Transformer-BiFPN neural network" (patent application number 202410567741.1) discloses a sheep face image detection method based on YOLOv8-Swin-Transformer-BiFPN neural network, which is suitable for detecting sheep faces. However, the sheep breeding environment is complex and diverse, including factors such as lighting changes and mutual occlusion. The existing target detection model still has a lot of room for improvement in practical applications. By improving the existing model, a scientific, objective and efficient method for identifying the behavior of housed sheep is established. The daily behavior of sheep can be quickly and accurately identified, and their health status, growth and development stage, and possible problems can be judged. This is of great significance to the refined management of sheep farming, disease prevention, and improvement of animal welfare. Summary of the invention

[0005] In view of the defects in the prior art, the purpose of the present invention is to provide a method for identifying the behavior of housed sheep based on improved YOLOv8n, which can accurately and quickly detect the location of sheep and identify their behavior, and realize real-time detection of sheep behavior in real sheep farming scenarios. The problem of low accuracy in sheep behavior recognition in crowded conditions is solved.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is: a method for identifying behavior of housed sheep based on improved YOLOv8n, comprising the following steps:

[0007] Step S1. Acquire the behavior video data of housed sheep;

[0008] Step S2. Preprocess the video data to construct a sheep image database;

[0009] Step S3. Based on the image of the shed sheep, use a labeling tool to label the sheep behavior;

[0010] Step S4. Divide the processed images into a training set, a validation set and a test set, and use data augmentation to expand the data diversity of the training set;

[0011] Step S5. Use depthwise separable convolutions to replace all standard convolutions in the CSPDarkNet53 feature extraction network and embed the proposed PCBAM attention module into the neck network of YOLOv8n;

[0012] Step S6. Input the expanded training set into the improved YOLOv8n network model for training; iterate using the CioU loss function to obtain the optimal model.

[0013] Based on the above technical solution, step S1 specifically includes:

[0014] Step S11. Install a high-definition surveillance camera diagonally above the sheep house;

[0015] Step S12. Use a high-definition surveillance camera to continuously record the daily behavior videos of sheep in a house environment to ensure that multiple time periods of sheep activities, different lighting conditions and diverse behavior scenes are covered to obtain original video data.

[0016] On the basis of the above technical solution, step S2 specifically includes:

[0017] Step S21. Screening the collected original videos to select videos with more sheep activities;

[0018] Step S22: Using video frame segmentation technology, obtain video images and build a database of images of housed sheep.

[0019] Based on the above technical solution, step S3 specifically includes:

[0020] Step S31. Use the open source software LabelImg image annotation tool to annotate the daily behavior of sheep in the video image and label each rectangular box;

[0021] Step S32: Each annotation will generate an annotation file in .XML format, and the generated .XML file will be converted into .txt format.

[0022] On the basis of the above technical solution, in step S31, three daily behaviors of sheep in the image are labeled as "lie", "stand" and "eat", corresponding to the three behaviors of lying, standing and eating, respectively.

[0023] Based on the above technical solution, step S4 specifically includes:

[0024] Step S41. Divide the dataset images and the corresponding .txt format annotation files into training set, validation set and test set in a ratio of 6:2:2, which are used for training, validation and testing of subsequent models respectively;

[0025] Step S42: Use data augmentation to expand the training set and increase the diversity of the data.

[0026] On the basis of the above technical solution, the data augmentation methods include left-right flipping, up-down flipping, scaling, translation, irregular cropping, size transformation, angle transformation and color gamut transformation.

[0027] Based on the above technical solution, step S5 specifically includes:

[0028] Step S51. Use a depthwise separable convolutional module to replace the standard convolutional module in the backbone network;

[0029] Step S52. Design a PCBAM module, embed the PCBAM attention module into the neck network, and improve the feature fusion capability of the neck network. The PCBAM attention module is divided into a channel attention module and a spatial attention module.

[0030] Step S53. The input original feature map is subjected to maximum pooling and average pooling operations to obtain a maximum pooling vector and a spatial pooling vector, and two vectors of the same latitude are obtained after passing through a shared fully connected layer. The two vectors are added together and then passed through a sigmoid activation function to finally obtain the channel attention weight. The channel attention module expression is:

[0031]

[0032] Among them, F represents the original input feature map; M C (F) represents the channel attention weight; MaxPool is the global maximum pooling; AvgPool is the global average pooling; MLP is the shared fully connected layer; W1 and W0 are the weights of the fully connected layer; and They are the average pooling vector and the maximum pooling vector respectively;

[0033] Step S54. The original feature map is input into the spatial attention module, and the maximum pooling operation and the average pooling operation are performed respectively to obtain two 1×H×W feature maps, which are concatenated to obtain a 2×H×W feature map. The feature map passes through a convolution layer with a convolution kernel size of 7×7 and then passes through the sigmoid function to finally obtain the spatial attention weight. The spatial attention module expression is:

[0034]

[0035] Among them, M S (F) represents the spatial attention weight; f 7×7 Indicates a convolution operation with a convolution kernel size of 7×7; is the average pooling vector; is the maximum pooling vector;

[0036] Step S55. Weight the output channel attention weights and spatial attention weights with the original input features to obtain the final output feature map.

[0037] Based on the above technical solution, the calculation formula in step S55 is:

[0038]

[0039] Among them, F ′ is the final output feature map.

[0040] Based on the above technical solution, step S6 specifically includes:

[0041] Step S61: input the training set after data expansion processing into the improved YOLOv8n network model for training;

[0042] Step S62. Use the CioU loss function to optimize the border prediction ability of the model. During the training process, the model is continuously iterated, and the model parameters are adjusted through the back propagation algorithm and the gradient descent method to minimize the loss function and obtain the optimal model.

[0043] The beneficial effects of the present invention are:

[0044] The present invention improves the YOLOv8n network model for the specific task of sheep behavior recognition and detection in housed scenarios, including adding an attention module and replacing the standard convolution in the feature extraction network with a deep separable convolution. The improved YOLOv8n model can accurately detect sheep behavior in complex backgrounds, crowded and occluded sheep, reduce the probability of missed detection and false detection, and improve the model detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is an example diagram of image enhancement of the sheep behavior recognition dataset of the present invention;

[0046] Figure 2 It is a network model structure diagram of the present invention;

[0047] Figure 3 It is a structural diagram of the PCBAM module of the present invention;

[0048] Figure 4 It is a detection effect diagram of the present invention. DETAILED DESCRIPTION

[0049] Embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions.

[0050] The following is a further description of the specific embodiments of the present invention in conjunction with the drawings of the specification, so that the technical solution and its beneficial effects of the present invention are clearer and more explicit. The following description of the embodiments with reference to the drawings is exemplary and intended to explain the present invention, but cannot be understood as limiting the present invention.

[0051] The embodiment of the present invention provides a method for identifying behavior of housed sheep based on improved YOLOv8n, comprising the following steps:

[0052] Step S1. Obtaining the behavior video data of the housed sheep; specifically, step S1 specifically includes:

[0053] Step S11. Install a high-definition surveillance camera diagonally above the sheep house;

[0054] Step S12. Use a high-definition surveillance camera to continuously record the daily behavior videos of sheep in a house environment to ensure that multiple time periods of sheep activities, different lighting conditions and diverse behavior scenes are covered to obtain original video data.

[0055] Step S2. Preprocess the video data to construct a sheep image database; specifically, step S2 includes:

[0056] Step S21. Screening the collected original videos to select videos with more sheep activities;

[0057] Step S22: Using video frame segmentation technology, obtain video images and build a database of images of housed sheep.

[0058] Step S3. Based on the image of the shed sheep, use a labeling tool to label the sheep behavior; specifically, step S3 specifically includes:

[0059] Step S31. Use the open source software LabelImg image annotation tool to annotate the daily behaviors of sheep in the video image and label each rectangular box; in this embodiment, the three daily behaviors of sheep in the image are labeled as "lie", "stand" and "eat", corresponding to the three behaviors of lying, standing and eating, respectively.

[0060] Step S32: Each annotation will generate an annotation file in .XML format, and the generated .XML file will be converted into .txt format.

[0061] Step S4. Divide the processed images into a training set, a validation set and a test set, and use data augmentation to expand the data diversity of the training set; specifically, step S4 specifically includes:

[0062] Step S41. Divide the dataset images and the corresponding .txt format annotation files into training set, validation set and test set in a ratio of 6:2:2, which are used for training, validation and testing of subsequent models respectively;

[0063] Step S42: Use data augmentation to expand the training set and increase the diversity of the data.

[0064] Specifically, the data augmentation methods include left-right flipping, up-down flipping, scaling, translation, irregular cropping, size transformation, angle transformation, and color gamut transformation.

[0065] Step S5. Use depthwise separable convolution to replace all standard convolutions in the CSPDarkNet53 feature extraction network, and embed the proposed PCBAM attention module into the neck network of YOLOv8n; specifically, step S5 specifically includes:

[0066] Step S51. Use a depthwise separable convolutional module to replace the standard convolutional module in the backbone network;

[0067] Step S52. Design a PCBAM module, embed the PCBAM attention module into the neck network, and improve the feature fusion capability of the neck network. The PCBAM attention module is divided into a channel attention module and a spatial attention module.

[0068] Step S53. The input original feature map is subjected to maximum pooling and average pooling operations to obtain a maximum pooling vector and a spatial pooling vector, and two vectors of the same latitude are obtained after passing through a shared fully connected layer. The two vectors are added together and then passed through a sigmoid activation function to finally obtain the channel attention weight. The channel attention module expression is:

[0069]

[0070] Among them, F represents the original input feature map; M C (F) represents the channel attention weight; MaxPool is the global maximum pooling; AvgPool is the global average pooling; MLP is the shared fully connected layer; W1 and W0 are the weights of the fully connected layer; and They are the average pooling vector and the maximum pooling vector respectively;

[0071] Step S54. The original feature map is input into the spatial attention module, and the maximum pooling operation and the average pooling operation are performed respectively to obtain two 1×H×W feature maps, which are concatenated to obtain a 2×H×W feature map. The feature map passes through a convolution layer with a convolution kernel size of 7×7 and then passes through the sigmoid function to finally obtain the spatial attention weight. The spatial attention module expression is:

[0072]

[0073] Among them, M S (F) represents the spatial attention weight; f 7×7 Indicates a convolution operation with a convolution kernel size of 7×7; is the average pooling vector; is the maximum pooling vector.

[0074] Step S55. Weight the output channel attention weights and spatial attention weights with the original input features to obtain the final output feature map.

[0075] Specifically, the calculation formula in step S55 is:

[0076]

[0077] Among them, F ′ is the final output feature map.

[0078] Step S6: Input the expanded training set into the improved YOLOv8n network model for training; iterate using the CioU loss function to obtain the optimal model. Specifically, step S6 includes:

[0079] Step S61: input the training set after data expansion processing into the improved YOLOv8n network model for training;

[0080] Step S62. Use the CioU loss function to optimize the border prediction ability of the model. During the training process, the model is continuously iterated, and the model parameters are adjusted through the back propagation algorithm and the gradient descent method to minimize the loss function and obtain the optimal model.

[0081] The present invention will be further described below by means of specific examples.

[0082] S101 obtains video data of daily behavior of housed sheep: uses high-definition surveillance cameras to continuously record videos of daily behavior of sheep in a house environment, ensuring coverage of multiple time periods of sheep activities, different lighting conditions, and diverse behavioral scenarios, to obtain original video data.

[0083] Furthermore, the specific location of the venue is not limited in the embodiments of the present application, and can be selected according to the actual situation. For example, the present application selects Qinyang City, Jiaozuo City, Henan Province as the data collection point. Preferably, the camera pixel should meet 10 million pixels or more, and the camera installation position should be selected in the time period of 9:00-11:30 in the morning and 2:30-5:00 in the afternoon with sufficient light and soft sunlight to ensure that the collected image is clear enough and prevent the image from being overexposed.

[0084] S201 preprocesses the above video data to obtain a sheep behavior dataset: first, the collected original videos are screened, and videos with more sheep activities are selected, and video images are obtained using video framing technology. Images with poor quality, such as images with severe exposure and severe blur, are eliminated to prevent images with poor quality from having a negative impact on the network. After screening, 1,185 images are obtained.

[0085] S301 uses professional image annotation tools to label the three daily behaviors of sheep in the image as "lie", "stand" and "eat", corresponding to the three behaviors of lying down, standing and eating respectively.

[0086] S401 divides the processed images into a training set, a validation set and a test set, and uses data augmentation to expand the data diversity of the training set.

[0087] Furthermore, the labeled images were divided into training set, validation set and test set in a ratio of 6:2:2, and finally a dataset of daily behavior of housed sheep was obtained.

[0088] Furthermore, we use data augmentation methods such as left-right flipping, up-down flipping, and scaling to expand the training set to increase the diversity of the data, and finally expand the training set to 4 times the original size. Figure 1 shown.

[0089] S501 improves the YOLOv8n model by replacing all standard convolutions in the CSPDarkNet53 feature extraction network with deep separable convolutions, and embedding the proposed PCBAM attention module into the neck network of YOLOv8n to reduce model complexity and parameter quantity, enhance the model's selectivity for key features, and improve the model's generalization ability and computational efficiency. The improved model structure is as follows: Figure 2 shown.

[0090] Furthermore, depthwise separable convolution is used to replace the standard convolution in the backbone network. The standard convolution extracts spatial features and channel features at the same time, while the depthwise separable convolution decomposes the feature extraction into two processes. First, a convolution kernel is applied to each channel of the input for convolution without changing the number of channels. Then, the number of channels is adjusted for the result of the channel-by-channel convolution. The computational complexity of the depthwise separable convolution is much smaller than that of the standard convolution.

[0091] Furthermore, the PCBAM attention module is embedded in the neck network of YOLOv8n, which significantly improves the model's ability to identify and utilize key features. The effect is more obvious when dealing with complex backgrounds and resource-constrained scenes, while enhancing the generalization ability and interpretability of the model. After multiple sets of experimental tests, the PCBAM attention module is embedded in the P5 layer of the neck network, and the network feature extraction ability is the best. The PCBAM module integrates the channel attention module and the spatial attention module, combining these two sub-modules in parallel. The PCBAM module is as follows Figure 3 shown.

[0092] The formula of the channel attention module is as follows:

[0093]

[0094] The formula of the spatial attention module is as follows:

[0095]

[0096] The formula of PCBAM module is as follows:

[0097]

[0098] S601 inputs the expanded training set into the improved YOLOv8n network model for training; it iterates using the CioU loss function to obtain the optimal model.

[0099] Furthermore, configure the operating environment of the PD-YOLO network model.

[0100] For example, the Ubuntu 18.04.5 operating system is used, the CPU is Intel Xeon Processor, the CUAD version is 11.7, the deep learning framework is Pytorch1.13.1, the compilation environment is Python3.7, and the NVIDIA GeForceRTX 3060 graphics card is used for training. Various environments are installed according to the configuration files in the YOLO source code.

[0101] Furthermore, the sheep behavior dataset mentioned in S401 is used to configure the data script Sheep.yaml of the PD-YOLO network model to ensure that the network model reads the specified data.

[0102] Furthermore, configure the network model hyperparameters, such as the number of batch samples, number of iterations, learning rate, and optimizer.

[0103] Exemplarily, the number of batch samples used in this application is 16, the number of iterations is 300, the learning rate is 0.01, and SGD is used as the optimizer.

[0104] Input the training set images mentioned in S401 into the network to start training, obtain the prediction information in the optimal network, then use Soft-NMS and set thresholds to filter out the detection boxes that meet the conditions from the prediction information, draw the final detection box on the sheep picture, and detect the predicted category information of the detection box in the upper left corner area of ​​the detection box, so as to complete the recognition of different sheep behaviors.

[0105] The detection effect of the present invention is as follows Figure 4 Compared with the prior art, the PD-YOLO model proposed in the present invention can accurately detect sheep behavior and achieve a higher detection speed, meeting the real-time detection requirements in industrial applications and providing an effective technical support for smart sheep farming.

[0106] In the description of the specification, the description with reference to the terms "one embodiment", "preferably", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention, and the schematic expressions of the above terms in this specification do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0107] The present invention is not limited to the above-mentioned embodiments. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention. The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A method for identifying behavior of housed sheep based on improved YOLOv8n, characterized in that: The following steps are involved: Step S1. Acquire the behavior video data of housed sheep; Step S2. Preprocess the video data to construct a sheep image database; Step S3. Based on the image of the shed sheep, use a labeling tool to label the sheep behavior; Step S4. Divide the processed images into a training set, a validation set and a test set, and use data augmentation to expand the data diversity of the training set; Step S5. Use depthwise separable convolutions to replace all standard convolutions in the CSPDarkNet53 feature extraction network and embed the proposed PCBAM attention module into the neck network of YOLOv8n; Step S6. Input the expanded training set into the improved YOLOv8n network model for training; iterate using the CioU loss function to obtain the optimal model.

2. The method for identifying behavior of housed sheep based on improved YOLOv8n as claimed in claim 1, characterized in that: Step S1 specifically includes: Step S11. Install a high-definition surveillance camera diagonally above the sheep house; Step S12. Use a high-definition surveillance camera to continuously record the daily behavior videos of sheep in a house environment to ensure that multiple time periods of sheep activities, different lighting conditions and diverse behavior scenes are covered to obtain original video data.

3. The method for identifying behavior of housed sheep based on improved YOLOv8n as claimed in claim 1, characterized in that: Step S2 specifically includes: Step S21. Screening the collected original videos to select videos with more sheep activities; Step S22: Using video frame segmentation technology, obtain video images and build a database of images of housed sheep.

4. The method for identifying behavior of housed sheep based on improved YOLOv8n as claimed in claim 1, characterized in that: Step S3 specifically includes: Step S31. Use the open source software LabelImg image annotation tool to annotate the daily behavior of sheep in the video image and label each rectangular box; Step S32: Each annotation will generate an annotation file in .XML format, and the generated .XML file will be converted into .txt format.

5. The method for identifying behavior of housed sheep based on improved YOLOv8n as claimed in claim 4, characterized in that: In step S31, the three daily behaviors of sheep in the image are labeled as "lie", "stand" and "eat", corresponding to the three behaviors of lying, standing and eating, respectively.

6. The method for identifying behavior of housed sheep based on improved YOLOv8n as claimed in claim 4, characterized in that: Step S4 specifically includes: Step S41. Divide the dataset images and the corresponding .txt format annotation files into training set, validation set and test set in a ratio of 6:2:2, which are used for training, validation and testing of subsequent models respectively; Step S42: Use data augmentation to expand the training set and increase the diversity of the data.

7. The method for identifying behavior of housed sheep based on improved YOLOv8n as claimed in claim 6, characterized in that: The data augmentation methods include left-right flipping, up-down flipping, scaling, translation, irregular cropping, size transformation, angle transformation and color gamut transformation.

8. The method for identifying behavior of housed sheep based on improved YOLOv8n as claimed in claim 1, characterized in that: Step S5 specifically includes: Step S51. Use a depthwise separable convolutional module to replace the standard convolutional module in the backbone network; Step S52. Design a PCBAM module, embed the PCBAM attention module into the neck network, and improve the feature fusion capability of the neck network. The PCBAM attention module is divided into a channel attention module and a spatial attention module. Step S53. The input original feature map is subjected to maximum pooling and average pooling operations to obtain a maximum pooling vector and a spatial pooling vector, and two vectors of the same latitude are obtained after passing through a shared fully connected layer. The two vectors are added together and then passed through a sigmoid activation function to finally obtain the channel attention weight. The channel attention module expression is: Among them, F represents the original input feature map; M C (F) represents the channel attention weight; MaxPool is the global maximum pooling; AvgPool is the global average pooling; MLP is the shared fully connected layer; W1 and W0 are the weights of the fully connected layer; and They are the average pooling vector and the maximum pooling vector respectively; Step S54. The original feature map is input into the spatial attention module, and the maximum pooling operation and the average pooling operation are performed respectively to obtain two 1×H×W feature maps, which are concatenated to obtain a 2×H×W feature map. The feature map passes through a convolution layer with a convolution kernel size of 7×7 and then passes through the sigmoid function to finally obtain the spatial attention weight. The spatial attention module expression is: Among them, M S (F) represents the spatial attention weight; f 7×7 Indicates a convolution operation with a convolution kernel size of 7×7; is the average pooling vector; is the maximum pooling vector; Step S55. Weight the output channel attention weights and spatial attention weights with the original input features to obtain the final output feature map.

9. The method for identifying behavior of housed sheep based on improved YOLOv8n as claimed in claim 8, characterized in that: The calculation formula in step S55 is: Among them, F ′ is the final output feature map.

10. The method for identifying behavior of housed sheep based on improved YOLOv8n according to claim 1, characterized in that: Step S6 specifically includes: Step S61: input the training set after data expansion processing into the improved YOLOv8n network model for training; Step S62. Use the CioU loss function to optimize the border prediction ability of the model. During the training process, the model is continuously iterated, and the model parameters are adjusted through the back propagation algorithm and the gradient descent method to minimize the loss function and obtain the optimal model.

Images