DCPNet: Distribution Calibration Prototypical Network for Few-Shot Image Classification

Deep learning has witnessed significant advancements in various tasks and has displayed exceptional performance. However, traditional deep learning techniques often necessitate the utilization of extensive labeled data for training, a requirement that is challenging to fulfill in many real-world scenarios. This limitation has given rise to the field of few-shot learning (FSL). In this paper, we introduce a Distribution Calibration Prototypical Network (DCPNet), aiming to address the limitations of prototypical networks in terms of their weak feature extraction capabilities and the inability of their classifier boundaries to align with the dataset. DCPNet incorporates a parallel hierarchical feature extraction module and a few-shot differentiation loss function to fine-tune the metric learning for better feature representation. This approach employs a parallel approach to extract features based on the semantic depth of image hierarchical extraction and incorporates contrastive learning to achieve feature vector fusion. Furthermore, DCPNet incorporates an improved distribution calibration method that leverages information from the base class dataset to align classifier boundaries with the dataset. To validate our approach, we conducted comparative experiments on datasets such as Mini-Imagenet, Omniglot, and CUB using classical baseline methods. In additional, we conducted ablation experiments on the Mini-Imagenet to assess the performance effectiveness of each component of the model. The results demonstrate that the proposed method presented in this paper outperforms other approaches and offer new insights into the field of few-shot image classification.