A pig face recognition method for distinguishing features
Wang Shi-gang, Wang Jian, Chen Meimei, Jinyang Wu
Abstract
Convolutional neural networks (CNNs) have been widely used in the field of computer vision, greatly improving their technical level. In the existing network neural networks, most of them use Softmax loss function as the supervision signal to train the deep model. In order to improve the recognition ability of deep learning features, this paper proposes a new recognition supervision loss function-a combination of center loss and L-Softmax (Center-L-Softmax) loss function. Specifically, the center loss learns the center of each type of deep feature, and penalizes the distance between the deep feature and its corresponding class center. More importantly, we proved that the proposed central loss function is trainable and easy to optimize in neural networks. Under the joint supervision of L-Softmax loss and center loss, we can train a robust ResNet to obtain as many deep features as possible with two key learning objectives, namely the loss between classes and the loss within classes. Compactness, which is very important for pig face recognition. Our RseNet (under such joint supervision) shows advanced accuracy on the data set (provided by insurance companies), significantly improving previous pig face recognition results and verification tasks.