DSE- ResNet: Dynamic Channel Attention for Lightweight Visual Recognition

Deep learning-driven image classification faces significant challenges in small-sample scenarios like medical imaging and rare species recognition. This paper proposes a novel few-shot learning method integrating dynamic channel attention with optimized residual networks. Our key innovation is an adaptive compression ratio mechanism that automatically adjusts attention-layer dimensionality reduction based on sample quantity and feature complexity. Combined with enhanced residual blocks, the framework maintains gradient stability while improving feature discrimination. Theoretical and experimental analyses demonstrate that the co-optimization of attention weights and residual pathways achieves optimal feature selection-retention balance. Evaluations on CIFAR-10 show state-of-the-art performance, attaining 82.3 % accuracy in 5-way 1-shot tasks-a 12.4% improvement over meta-learning baselines. This approach effectively reduces low-dimensional information loss by 15.6% compared to static attention models, demonstrating strong potential for medical diagnostic applications requiring few-shot adaptation.