TailSafe: A Pig Head-to-Rear Contact Detection System Using Convolutional Neural Networks

Pig tail biting poses significant challenges in pig farm monitoring, serving as an indicator of underlying pen issues. Existing monitoring methods are limited in their scalability and invasiveness. The goal of the study is to ascertain the capability of the YOLOv5 (You Only Look Once) algorithm with CSPDarknet53 as the backbone in detecting pig heads and pig rears. An exploratory approach to determine the optimal YOLO model training hyperparameters for detection (image size, batch size, epoch) and for interaction (image size and Intersection over Union (IoU) thresholds) was utilized. The detected pig parts will be utilized to determine the presence of head-to-rear contact in pig pen images using the computation of the part's IoU compared to a calibrated interaction threshold. The topmost performing detection model was obtained by the configuration of 600×600 input size, batch size 8 and epoch 100 with highest mean average precision (mAP) at 0.930, 0.929 precision and 0.895 recall. Likewise, it was also the topmost performing model to ascertain interaction with input image size of 500×500 and IoU threshold of 0.05 in the interaction dataset. This configuration generated 76% accuracy, 90% recall, 62% precision, 73% F1-score, and 0.776 normalized Matthews Correlation Coefficient. This study introduces TailSafe, a simple decision support tool utilizing YOLOv5 and convolutional neural networks, enabling farmers to diagnose pig pen issues through potential tail biting outbreaks. TailSafe comprises two components: a detection method to identify pig heads and rears, and an interaction method to compute head-to-rear IoUs for contact identification.