Few-shot learning (FSL) presents critical limitations due to the scarcity of labeled training data, severely limiting models’ ability to learn discriminative feature representations. Issues, such as insufficient feature diversity and scale sensitivity, making it challenging for image classification tasks in FSL. To bridge this gap, we propose a novel adaptive multi-scale feature fusion (AMFF) network for FSL. AMFF includes two innovative modules, feature enrichment module (FEM) and feature fusion module (FFM). FEM increases feature richness through a multi-branch convolution structure and expands representation capacity with limited data. Then FFM reorganizes the multi-scale features with different weights, enabling the network to capture multi-scale information more effectively. Our approach systematically addresses the feature paucity problem through two complementary mechanisms: (1) expanding feature diversity via multi-scale feature extraction, and (2) enhancing feature discriminativeness through an adaptive feature fusion strategy. Extensive experiments on popular FSL benchmarks (miniImageNet, CUB) demonstrate our AMFF achieves competitive performances in few-shot classification tasks.