Local feature semantic alignment network for few-shot image classification

The goal of few-shot learning is to use a small number of labeled samples to train a machine learning model and then classify the unlabeled samples. Recent works, especially the methods based on image local feature representation in metric learning have achieved superior performance by utilizing the local invariant features and their rich discriminative information. However, the learned local features in the existing methods are not aligned when calculating their similarities, resulting in larger intra-class divergence and smaller inter-class divergence. In fact, the dominant object (local feature) of one image should only compare with the semantically relevant local feature of the other image. To address these issues, this paper proposes a few-shot learning approach (SANet) based on semantic alignment of local features. Specifically, we firstly obtain the local features of the query and support images by using a feature extraction module, and then compute the relation matrices of these local features. Using the above relation matrices, we respectively design an intra-class divergence rectification (intraDR) module and an inter-class divergence rectification (interDR) module to implement the local feature alignment and reduce the effect of the noise local features. The experimental results on multiple datasets show that, by aligning the local features, the proposed model can effectively minimize the intra-class divergence while maximizing the inter-class divergence, thus achieving better classification performance. The code for this paper can be accessed via https://github.com/SongQCode/SANet .