Zero-Shot Image Classification: Recent Status and Future Trends

Deep learning has achieved great success in image classification tasks. However, the majority of the methods can only recognize classes that have already appeared in the training set. In practical applications, it is also necessary to recognize classes that have not been appeared in the training set. Zero-shot learning can help to solve this problem. In zero-shot image classification, classes in the training set are entirely different from those used in the test set. This paper provides a systematic study on zero-shot image classification algorithms over the past few years. Firstly, we provide the research background of zero-shot image classification. Secondly, we introduce the basic concepts of various zero-shot image classification methods, including the semantic embedding and extraction of features. Thirdly, we categorize existing zero-shot image classification algorithms into three main groups - direct semantic prediction-based methods, space mapping-based methods, and generative model-based methods. We then analyze and compare their respective performance on three common datasets. Finally, we provide a summary of the technical challenges and future research directions of zero-shot image classification.