Feature Generating Network For Zero-Shot Classification

Abstract:
Due to the problem of mode collapse when reconstructing image features by generative adversarial networks, an improved zero-shot classification method for image feature generation is proposed. One discriminator is added to give the generated image high scores, and give the real samples low scores; at the same time, the semantic feature vector is mapped to the regularized visual feature space to alleviate the semantic gap problem. The image uses the ResNet-101 convolutional neural network model to extract the image feature vector, and finally the softmax classification loss function is used for the image classification. The experimental results show that compared with the original method, the accuracy of this method is increased by 0.9%, 1.6%, and 2.1% on the AWA, CUB, and FLO datasets, respectively.