Zero-shot Animal Recognition with Feature Generating Networks
Recent years, due to the breakthroughs in deep neural networks, deep learning has yielded significant performance on the recognition of supervised tasks in the fields of computer vision and machine learning. However, more and more new categories emerge in image recognition, it is extremely difficult to collect sufficient tagged training samples for each possible category which leads into an urgent question that how to enable the model to learn efficiently from a category with limited or none samples. Zero-shot learning is one way of addressing the lack of training samples (target domain) by learning from the visible data (source domain) of seen classes, but it could not improve recognition accuracy of unseen classes because only seen classes samples have been involved in the training. Most existing approaches fail to achieve a satisfactory result in zero-shot learning task. Therefore, in this paper, we propose a feature-generated network model based on the Generative Adversarial Network which could generate features automatically for zero-sample categories, and according to the features we could further learn the zero-sample classifier. This method improves the recognition accuracy by 3.7% on the AWA dataset and 1.3% on the CUB dataset. The experimental results show that our method achieves a good recognition effect.