Task‐oriented feature hallucination for few‐shot image classification

Abstract Data hallucination generates additional training examples for novel classes to alleviate the data scarcity problem in few‐shot learning (FSL). Existing hallucination‐based FSL methods normally train a general embedding model first by applying information extracted from base that have abundant data. In those methods, hallucinators are then built upon trained generate classes. However, these usually rely on general‐purpose embeddings, limiting their ability task‐oriented samples Recent studies shown task‐specific models, which adapted tasks, can achieve better classification performance. To improve performance of example is used proposed method perform generation. After initialization, hallucinator finetuned with guidance teacher–student mechanism. The contains two steps. An initial network and an dataset step. second step pseudo‐labelling process where pseudo‐labelled using support task fine‐tuning adjusted simultaneously. Both updated set knowledge distillation. experiments conducted four popular datasets. results demonstrate approach outperforms state‐of‐the‐art 0.8% 4.08% increases accuracy 5‐way 5‐shot tasks. It also achieves comparable 1‐shot