A multiscale semantic feature-based method for clothing image classification

To address the complexity of fashion images in pose, texture, and category distribution, we propose a hybrid architecture oriented toward multi-scale feature modeling and global semantic enhancement. By introducing a residual-connected multi-scale feature aggregation mechanism, the method effectively integrates low-level details with high-level semantics. On this basis, a global-dependency modeling module strengthens the representation of long-range relationships, thereby improving recognition performance on complex fashion images. Fashion-MNIST and DeepFashion experiments demonstrate the effectiveness of the proposed approach, achieving 95.6% Top-1 accuracy in Fashion-MNIST and 73.6% Top-1 accuracy in DeepFashion, surpassing existing methods.