Research And Implementation Of Lightweight Sheep Face Recognition Method Based On Attention Mechanism

In the production,operation and management of the modern large-scale sheep industry,sheep identification is a fundamental and important task to achieve precise breeding.In actual production applications,traditional identification methods based on RFID technology have disadvantages such as limited identification distance,high cost,and easy falling off of ear tags.In recent years,with the popularization of camera equipment in farms and the development of face recognition technology,the use of computer vision technology to realize individual identity recognition has been widely used,and many sheep face recognition models based on convolutional neural networks have achieved excellent results.However,many existing models have the problems of too many parameters and high computational complexity,making it difficult to deploy applications in the actual environment of sheep farms.Therefore,the construction of a lightweight sheep face recognition model has great application prospects and industrial value.This paper mainly starts from the essential principles of target detection and image recognition,in order to enable the model to be deployed and run on edge computing devices with limited computing power,taking into account the speed and accuracy of sheep face recognition tasks,conducted in-depth research and discussion on the construction method of lightweight sheep face detection and recognition model to improve the applicability of sheep face recognition model on edge computing devices.The main accomplishments of this research can be summarized as follows:(1)In order to achieve fast and accurate detection of sheep faces,a sheep face detection model is constructed based on the Retina Face model.The lightweight Mobile Net based on depthwise separable convolution is selected as the feature extraction network of the model; for the scene where sheep faces with different distances are detected in practical applications,an adaptive hole coding module is designed to enhance the feature receptive field and enhance the model’s ability to detect different distances.The detection ability of the scale sheep face;in order to further enhance the regression effect of the sheep face bounding box,the CIo U loss function is used as the sheep face bounding box loss function.The experimental results show that the average accuracy of the improved sheep face detection model reaches 97.12%,which is 3.54% higher than the original Retina Face model,the model size is reduced by 0.25 MB,and the computational complexity is reduced to 64.7%.(2)Research on the Construction Method of Sheep Face Recognition Model.Aiming at the problem that the similarity between sheep individuals is high and difficult to distinguish,a sheep face recognition model is constructed based on the channel mixing module in the Shuffle Net V2 lightweight model.Integrating the SKNet convolutional attention mechanism enhances the model’s ability to extract sheep face features of different scales; fuses the MPECANet hybrid pooling channel attention mechanism to enhance the feature extraction ability in the channel domain; the training samples for the sheep face recognition dataset are too small to be manually designed.To determine the optimal hyperparameters of the metric function,the adaptive cosine metric function(Adacos)is used to train the model to adaptively adjust to the optimal hyperparameters and speed up the model convergence.The experimental results show that the correct recognition rate of the improved sheep face recognition model reaches 91.52%,which is 4.67% higher than that of the Shuffle Net V2 model,the model size is reduced by 0.99 MB,and the calculation amount is only increased by 2MFLOPs.(3)Design and implementation of sheep face recognition system based on Jetson Nano.Aiming at the problem that the deep learning model is difficult to deploy in the actual environment of the sheep farm,the Jetson Nano edge computing device is used as the platform to deploy the sheep face detection and recognition model on the device,and the Py Qt-based framework is implemented on the device.sheep face recognition system.The video image collection of individual sheep in the sheep farm through the network camera realizes the realtime detection and recognition of sheep faces and visualizes the recognition results,which has strong practical significance for promoting the application of sheep face recognition.