Steel Surface Defect Detection Using an Enhanced YOLOv8s with PPA Mechanism and AFPN

Abstract

Accurate and real-time detection of steel surface defects is essential for industrial quality control, yet it remains challenging due to extreme scale variations and complex background noise. Conventional object detection models often suffer from feature degradation and spatial information conflicts when processing these microscopic and multi-scale defects. To overcome these limitations, we propose an enhanced YOLOv8s-based detector. First, a Paralleled Patch-Aware Attention (PPA) module is integrated to extract multi-scale defect features, adaptively emphasizing critical textures while suppressing irrelevant industrial background noise. Second, we optimize the neck architecture by introducing an Asymptotic Feature Pyramid Network (AFPN) equipped with an Adaptive Spatial Fusion (ASF) mechanism. This structure progressively fuses non-adjacent hierarchical features, effectively mitigating semantic gaps and spatial information conflicts during multi-level aggregation. Extensive experiments conducted on the NEU-DET dataset demonstrate that the proposed method significantly outperforms the baseline YOLOv8s and other state-of-the-art models. The enhanced model achieves a superior balance between detection precision (mAP@0.5) and real-time inference speed, making it highly suitable for practical industrial inspection tasks.