Research On Sheep Face Object Extraction And Identity Recognition Based On Image Segmentation

Sheep identification is the basic link for large-scale sheep farms to carry out intelligent farming.Currently,sheep farms mainly implement sheep identification through the installation of special livestock ear tags.In this identification method,sheep often bite and drop ear tags,and the installation of ear tags will cause stress response to sheep.To solve the above problems,this thesis uses the deep learning method in the field of image recognition to realize the identification of sheep through the image of sheep face.This is a non-contact identification method of sheep,which can avoid the stress reaction of sheep to install ear tags and improve the identification efficiency of sheep in sheep farms.The specific research work of this thesis is as follows:1.In order to quickly acquire a large number of sheep face images,a multi-angle sheep face image acquisition channel and software system were built.A camera with five angles was installed in the channel to collect 110 sheep face images,which were transmitted to the upper computer through the network.11,000 original sheep face images were selected and enhanced.Increased to a data set of 55,000 sheep face identification images.2.A sheep face segmentation and recognition method for Mask R-CNN model based on the mixed attention mechanism of pre-training network and CBAM is proposed to improve the accuracy of the model.Res Net-101 and data-enhanced images were selected for training after ablation test.The pre-training network was added,the improved I-CBAM mixed attention mechanism was added,and the regions with key feature information were improved while the model state was adjusted.The test showed that the segmentation effect was improved,the accuracy of front view was higher than that of left view and right view,and the detection accuracy of the total test set was increased by 1.32%.3.A sheep face segmentation and recognition method of Blend Mask model based on Res Next feature extraction and hyperparameter optimization design was proposed.Considering the need of deploying the model to the edge for verification,it was proposed to replace the Res Next feature extraction network based on Blend Mask network and carry out hyperparameter optimization at the same time.The training results show that the convergence effect of the improved model is improved,the m AP value of the model is increased by 0.9%,and the accuracy is increased by 1.43%.4.A sheep facial recognition verification system was deployed to guide the trained weight data into the verification system.The system tends to run smoothly after debugging.Therefore,this study can be applied to sheep identity management in farms.