| With the development of science and technology,the demand for intelligent management of animal husbandry has become increasingly obvious,and the use of scientific and technological means for individual animal identification has also become a research hotspot.Image recognition is one of the most widely used methods to recognize objects by extracting features from images.Since the accuracy of image feature extraction directly affects the recognition result,extracting the effective feature has become the most important step in animal identification.In this paper,after investigation,it is found that it is more difficult to carry out cattle eyes iris collection in real life,which leads to the decrease of network recognition accuracy.Therefore,this paper proposes a cattle eye recognition method based on feature extraction to provide a theoretical foundation and practical value for intelligent pastures.Specific contents include.1.Build image data acquisition system.Use OV5640 camera for image acquisition and transmit the collected images to the host computer via Ethernet,and es Tab.lish detection and feature extraction data sets.2.Target detection algorithm selection.A comparison experiment is conducted on the widely used Faster RCNN,YOLOv3 and YOLOv4 networks on the self-built cattle eye detection dataset,and YOLOv4 is selected as the target detection algorithm after analyzing the experimental results.3.An improved SqueezeNet cattle eye feature extraction algorithm is proposed for the text map with low network recognition accuracy.Firstly,a lightweight network SqueezeNet is built and the Mish function is used to replace the Re LU function in the original FIRE module to enhance the robustness of the network and speed up the feature extraction.And combined with Stochastic Pooling to improve the generalization ability of the network.Finally,the network optimization experiments are carried out on different Dropout parameters to obtain the optimal network model SqueezeNet-MS,which is verified on the self-built cattle eye image dataset and compared with several existing networks.4.An improved DenseNet cattle eye feature extraction algorithm is proposed to address the weak feature expression of the network.The DenseNet is constructed with two parts.Firstly,the shallow cattle eye features were put into each layer of the cattle eye image using dense connections to enhance the transmission of information within the network to obtain more features.Secondly,the SELU function is used to prevent overfitting of the network,and the extracted features are randomly discarded by Drop Block to obtain more accurate cattle eye image features and improve the recognition accuracy.The recognition accuracy of the improved DenseNet network is97.47%,and the recognition rate of the improved SqueezeNet network is the only96.59%,although the number of model parameters is 1.25×10~6 which is close to half of the number of parameters of the improved DenseNet network.The experimental results show that the improved networks all have high recognition of cattle eye features. |
