Research And Application Of Pig Individual Identification Method Based On Lightweight YOLOv4

With the continuous improvement of science and technology and the expansion of pig breeding scale,the traditional ear tag,breeder observation,tattooing and other individual pig identification methods exist to cause individual pig discomfort,high human cost,high tattooing error rate and other problems,which are difficult to meet the needs of large-scale pig farm breeding automation.Therefore,in a group breeding environment,it is of great practical significance to further improve the level of automated management of large-scale pig farms by using intelligent methods to automatically and effectively and quickly identify individual pigs,to realise the establishment of a file and card for them,and to establish a traceability platform for pig breeding.To address the above real-world problems,this paper learns to draw on similar non-invasive identification methods without harm,and proposes a pig identification model based on a lightweight YOLOv4 algorithm by way of recognizing the face of a pig with the help of convolutional neural network methods in the field of target detection.The main work in this paper is summarised as follows.(1)An experimental dataset for individual pig identification was constructed.In this paper,we obtained the original video data of 30 long white pigs from the public video footage provided by the JDD-2017 Jingdong Finance Global Data Explorer Competition,and went to the pig farm to obtain video footage of 25 fattening black pigs in the field.The two kinds of video data obtained are converted into image data,and through a series of pre-processing operations such as similarity de-weighting and image normalisation,and finally the data are manually annotated according to the experimental data format requirements to construct two experimental datasets for pig face detection.(2)A lightweight YOLOv4-based identification model for individual pigs is proposed.Firstly,the two datasets that have been processed are divided according to the training set,validation set and test set 6:2:2 respectively,and the two datasets are trained by constructing four mainstream target detection algorithms for each comparison experiment.The results show that the YOLOv4 algorithm performs best.Secondly,to address the problems of large size and slow detection of the original model,a lightweight model approach is proposed,which uses the lightweight network Mobile Net v3 for backbone feature network replacement,and uses depth-separable convolution in the network to replace the normal convolution.The improved model is 1/4 of the size of the original YOLOv4 model and has a memory footprint of only 54.2M,making it suitable for mobile devices.(3)A lightweight YOLOv4 verification platform for individual pig identification methods was designed.Using the lightweight YOLOv4 individual pig identification model proposed in this paper as the core,the trained model was developed using a front-and back-end separation model,deployed on the mobile platform using the Flask framework,and the detection and recognition performance of the model was tested and validated on two datasets respectively.The validation results show that the improved model has a detection speed of 45ms/sheet,an increase in detection accuracy of 1.99% and a maximum accuracy of 98% for the identification of individual pigs,and can complete the construction of pig breeding information in this platform and carry out real-time reporting of pig breeding information.