SCNNnet: Self-Attention Enhancement and Cross-Correlation Nearest Neighbor Network for Few-Shot Fine-Grained Social Visual Recognition

Few-shot learning (FSL) aims to extract discriminative features from only a few labeled samples and generalize them to unseen samples, which is a crucial capability for intelligent systems interacting in dynamic social environments. A promising research approach is to identify the most similar class by comparing the local descriptors between the input image and the category. However, previous methods often overlook the relevance and importance of local descriptors, which is particularly critical in fine-grained visual recognition, where subtle interclass differences require the model to accurately localize discriminative regions and evaluate the significance of each local descriptor. To address this issue, we propose aSelf-attention enhancement module (SAEM) and aCross-correlationNearestNeighbor neural network (SCNNnet). The SAEM assigns weights to each local descriptor, enabling the model to concentrate on more discriminative regions. The cross-correlation module (CCM) calculates relationships between image representations by generating a similarity matrix, which enhances regions that are highly similar to the target features. Extensive experiments on three few-shot fine-grained benchmark datasets demonstrate that SCNNnet achieves state-of-the-art performance, while offering insights into attention-driven decision processes-advancing the development of socially aware vision.