Contrastive Clustering Learning for Few-shot Image Classification

The objective of Unsupervised Few-Shot Learning (UFSL) is to obtain high-level representations using a limited amount of unlabeled data. However, in the actual process of feature extraction, the model merely focuses on a single focal point, failing to adequately exploit the rich information within the samples. In order to solve this issue, we propose a framework of Multi-Focal Semantic Access (MFSA). Our approach, during the training process, leverages both contrastive learning and deep clustering perspectives to better identify intra-instance and inter-instance semantic disparities, thus capturing multiple focal points of the samples. Specifically, contrastive learning generates instance-level prototypes to emphasize inter-instance differences, thereby enhancing the generalization capability of representations. Deep clustering uses random partitions to dynamically generate powerful clustering-level prototypes to focus on intra-instance differences, effectively mitigating the loss of intra-instance semantic information. Through the joint action of contrastive learning and deep clustering, MFSA effectively enhances the ability of extracting comprehensive representations for fuzzy scene tasks. Extensive experiments conducted on the miniImageNet dataset have demonstrated the superiority of MFSA. This is because, by focusing on multiple focal point of images to extract comprehensive representations, can effectively classify images in complex scenes.