Efficient Feature Extraction with Multi-Scale Attention Mechanism: A Lightweight Deep Learning Framework for Image Classification

Deep learning-based image classification has achieved remarkable progress in recent years, but the contradiction between model performance and computational efficiency remains a critical challenge for edge-device deployment. To address this issue, this paper proposes a lightweight deep learning framework integrated with an efficient multi-scale attention (EMA) module for high-performance feature extraction. The EMA module adopts a channel-grouping strategy and parallel multi-branch architecture to capture multi-scale contextual information without dimensionality reduction, which effectively avoids the loss of feature details caused by traditional attention mechanisms. Specifically, it divides input features into multiple subgroups and employs 1 × 1 and 3 × 3 convolutional branches to model local and global dependencies respectively, followed by cross-spatial learning to fuse complementary features across branches. The proposed framework is evaluated on three benchmark datasets (CIFAR-100, ImageNet-1k, and Tiny-ImageNet) against state-of-the-art lightweight models and attention mechanisms. Experimental results demonstrate that the proposed framework achieves a better trade-off between classification accuracy and computational cost. The proposed framework provides a promising solution for efficient image classification in resource-constrained scenarios.