FLYOLOv3 deep learning for key parts of dairy cow body detection

Abstract Accurate detection of the key parts of dairy cows is important to the intelligent behavior perception of cows. Owing to the complex background of dairy farms, the object characteristics of dairy cows are not obvious, and traditional object detection algorithms cannot detect the key parts with high precision. In this study, a deep learning network named FLYOLOv3 (FilterLayer YOLOv3) based on FilterLayer was tested to achieve the detection of key parts of dairy cows in complex scenes. Since images are unstable during the training process and initialization, particle noise was generated in feature maps after convolution. Therefore, the mean filtering algorithm was carried out, and a leaky rectifier function (Leaky ReLU) was applied to integrate the custom FilterLayer layer to reduce training interference. The artificial annotation was used firstly to mark the borders of the cow's head, back and legs in 1000 cow images, and then the FLYOLOv3 network was trained with the labeled samples. Finally, in order to verify the effectiveness of the algorithm, 500 images were randomly selected from 1000 cow images as the original training images, and the remaining images were used as test images. The proposed method was compared with the Faster R-CNN and YOLOv3 algorithm by using indicators such as accuracy, recall rate, average frame rate, and average accuracy. Test results showed that the accuracy of FLYOLOv3 algorithm was 99.18%, the recall rate was 97.51%, the average frame rate was 21 f/s, and the average precision was 93.73%. All the results showed that the proposed algorithm could be used for high-precision detection of the key parts of dairy cows.