Cow identification based on fusion of deep parts features

Livestock identification is of great significance for achieving precision livestock farming as it is a prerequisite of modern livestock management and automatic behaviour analysis. With respect to cow identification, methods based on computer vision have been widely considered due to their advantage of non-contact and practicality. In this paper, a novel non-contact cow identification method based on fusion of deep parts features is proposed. First, a set of side-view images of cows were captured, and then the YOLO object detection model was applied to locate the cow object in each original image, which was then divided into three parts, head, trunk and legs, by a part segmentation algorithm using frame differencing and segmentation span analysis. Then, three independent convolutional neural networks (CNNs) were trained to extract deep features from these three parts, and a feature fusion strategy was designed to fuse the features, i.e., deep parts feature fusion. Finally, a support vector machine (SVM) classifier trained by the fused features was used to identify each individual cow. The proposed method achieved 98.36% cow identification accuracy on a dataset containing side-view images of 93 cows, which outperformed existing works. Experimental results showed the effectiveness of the proposed cow identification method and the good potential for this method in individual identification of other livestock.