Mixture Loss Function-based Classification Network for Few-shot Learning

Data augmentation technology can effectively solve the problem of few-shot image classification, but many approaches based on data augmentation generated sample have poor discriminability, which negatively affects the classifier and classification accuracy needs to be improved. In this regard, based on the TFFCN, a mixture loss function-based classification network for few-shot learning (MLFCN) is proposed. First of all, for the prototype tensor features with a simple mean representation, ignoring the different contributions of the sample features of support set for per class to the class prototype tensor features, so that the generated prototype tensor features can not well represent sample features of per class, affecting the discriminability of the generated samples. Introduce the improved adaptively spatial feature fusion module, it uses the convolution, considers the support set of different sample features on the prototype tensor, adaptively assign weights, then weighted and summed to obtain a more representative prototype tensor feature; secondly, aiming at the blurred classification boundary of similar categories in the generated sample features, a mixture loss function is proposed. The loss function is to add the inter-class mean squared error term to the mean square error loss, and stretches the inter-class distance of the generated tensor feature, so that the classification boundary is clearer, so as to train a better classifier. Experimental results show that the improved adaptively spatial feature fusion and the mixing loss function can augment the discriminability of the generated tensor features and further improve the classification accuracy.