Teat detection of dairy cows based on deep learning neural network FS-YOLOv4 model

Abstract

To realize automatic milking operation of large-scale dairy cow breeding, an efficient detection method for dairy cow teats, based on the improved deep learning neural network Fourth feature scale-spatial pyramid pooling-You Only Look Once v4 (FS-YOLOv4) model, was proposed to solve the problems of poor accuracy and detection slowly in the complex milking parlour in this paper. In the backbone network of the YOLOv4 model, the CSPDarknet module which had fewer parameters and low complexity was replaced by the CSResNeXt module. To improve the feature extraction of the feature map, the detection feature scale, and the Spatial Pyramid Pooling structure were added to the path augmentation network. To evaluate the performance of the FS-YOLOv4 teat detection, the test set containing 200 images was used to test the model. The results showed that the precision, recall, mAP value, and F1-score were 98.81%, 98.46%, 98.26%, and 98.63% respectively. The results were further compared with those of single-stage the YOLOv4 model, the YOLOv5 model and two-stage the Faster RCNN model. It was verified that the performance of the FS-YOLOv4 model significantly improved the detection accuracy and detection time of dairy cow's teats, and the anti-noise ability was significantly enhanced. This study provides useful exploration for automated milking.