Competing ratio loss for discriminative multi-class image classification

The development of deep convolutional neural network architecture is critical to the improvement image classification task performance. Many studies use and focus on modifying structure improve Conversely, our study focuses loss function design. Cross-entropy Loss (CEL) has been widely used for training multi-class classification. Although CEL successfully implemented in several tasks, it only posterior probability correct class. For this reason, a negative log likelihood ratio (NLLR) was proposed better differentiate between class competing incorrect ones. However, during network, value NLLR not always positive or negative, which severely affects convergence NLLR. Our (CRL) calculates classes further enlarge difference classes. We added hyperparameters CRL, thereby ensuring its be that update size backpropagation suitable CRL’s fast convergence. To demonstrate performance we conducted experiments general tasks (CIFAR10/100, SVHN, ImageNet), fine-grained (CUB200–2011, Stanford Car), challenging face age estimation (using Adience). Experimental results showed effectiveness robustness different architectures tasks. Code released at https://github.com/guoyurong0104/CRL-code.