Lightweight Pig Face Detection Method Based on Improved YOLOv8

With the increasing scale of pig farming and the frequent occurrence of swine fever, effective management of farms and individual identification of pigs have become crucial challenges in the pig farming industry. Facial recognition of pigs has been widely researched in recent years, however, The effect of pig face recognition in practical application is not good. Collaborating with real pig farms, we found that the main problems were the lack of accuracy and processing speed in the detection of pig face images, the inaccuracy of prediction box, and the inability of low-power devices such as mobile phones and embedded devices to run large-scale models. To address these challenges, we proposes a lightweight pig face detection method based on improved YOLOv8. The method mainly uses YOLOv8 and incorporates an Inverted Residual Mobile Block (iRMB) to enhance the backbone network, striking a balance between lightweight design and detection accuracy. For the relatively complex dual-branch detection head in YOLOv8, we propose a channel-grouped multi-scale convolution method, Four-Level Convolution (FLConv), and propose a single-path shared parameter detection head, Shared Parameter head (SPHead), to reduce the computation of the model. To enhance prediction box accuracy, we replace the CIoU loss function with the MPDIoU loss function, which better improves the accuracy of bounding box regression. Our approach features a very small number of parameters and computations while achieving high accuracy and stability. Tested on a self-built pig face dataset, our method, compared to the most lightweight YOLOv8 model, YOLOv8n, while the mAP index increased by 0.2%, the number of parameters decreased by 17.0 % and the calculation volume decreased by 38.2 %.