Research On Deer Face Recognition Based On Deep Learning

In recent years,driven by the national strategy of industrial intelligent transformation and upgrading,all walks of life in society are accelerating the reform of traditional industries and seeking new growth points for the Chinese economy.Among them,intelligent aquaculture also goes hand in hand with the development strategy.With the employment of the Internet,big data and artificial intelligence technologies,the aquaculture industry is promoted to be scientific,intensive and precise.The development of intelligent farming has led to an increasing demand for animal identification,and at the same time,animal traceability is also driven by the animal’s need for quality control and welfare management.At present,the identification of deer is mainly realized by RFID technology,and this method has many disadvantages in practical application.On the one hand,electronic tags need to be attached to the surface of animals or embedded in the body,which will bring great pain to deer,and the easy loss of physical media such as ear tags increases management costs for deer farms.On the other hand,the technology itself is limited by factors such as distance and environment,which adds some complexity to the recognition.In view of the above factors,this paper uses a non-contact deep learning method to propose a deer face recognition method based on object detection,which can automatically extract high-dimensional abstract features in deer face images and register the deer identity,lays the foundation for further realization of deer information traceability and automatic monitoring.The work of this paper mainly includes the following three parts: First,the video data of deer faces are obtained by artificial field collection and make deer face datasets.To ensure the robustness of the model,the video data of 51 deer from different angles are collected under different lighting conditions,and the videos are broken down into frame-by-frame images,and then use the SSIM algorithm to clean the image,and finally build deer face datasets for detection model and recognition model.Secondly,YOLOv4 object detection algorithm is used to detect deer faces,and the input image is judged whether it contains deer face.If it does,the detection box of deer face needs to be given.According to the detection box,the image containing only the face area is cropped,and the interference factors such as the background of the deer house are removed,so that the model pays more attention to the face information.Finally,the DFRNet（Deer Face RecognitionNetwork）deer face recognition network is constructed,which is the core work of this paper.The DFRNet network takes ResNet18 as the main structure,and introduces the CC-SE attention mechanism combining channel attention and position attention to extract deer face features with enhanced representation.The CC-SE attention mechanism is a hybrid attention structure with channel dependencies and spatial location information correlation proposed in this paper.The influence of key point features such as eyes,nose and mouth of deer face and facial features such as hair and texture on recognition results and the correlation between deer facial features are considered comprehensively.To verify the effectiveness of the proposed DFRNet algorithm,experiments are performed on the original multi-angle deer face dataset and the enhanced multi-angle deer face dataset constructed by ourselves.The results show that the proposed model achieves 93.69% and 95% recognition accuracy on the test set of the original multiangle deer face dataset and the test set of the enhanced multi-angle deer face dataset,respectively.Compared with ResNet18,ResNet50,and ResNet101 networks as feature extraction networks,DFRNet has different degrees of improvement in recognition accuracy,false positive rate,and false negative rate.In this paper,the deer face is located by YOLOv4 object detection model,and based on the proposed CC-SE structure,the DFRNet deer face recognition model combined with the attention mechanism is constructed.The results of the detection model are cut out and input to the recognition network to complete the automatic detection and recognition of deer faces.The method proposed in this paper has a strong practical application background and has good adaptability to other animal facial recognition.