A YOLOv8-Based Clothing Detection Framework Incorporating Semantic Uncertainty Modeling

Abstract

Garment style recognition is critical for intelligent retail but suffers from semantic uncertainties caused by high intra-class similarity and non-rigid deformations. Existing detection and attention-based methods often apply uniform attention across all feature levels. This approach limits fine-grained discrimination in complex scenarios. To address this gap, this study proposes a hierarchical attention-enhanced YOLOv8 framework to optimize recognition precision. By integrating a Convolutional Block Attention Module (CBAM) at the mid-level and an Efficient Channel Attention (ECA) module at the high-level, the model effectively strengthens structural perception and suppresses semantic channel dispersion. Experimental results show that the proposed method achieves 84.99% mAP@0.5, an 11.82% improvement over the baseline, while maintaining real-time performance at 126 FPS. This framework improves fine-grained garment recognition and provides a practical solution for intelligent retail applications in complex scenarios.