Improved YOLOv8 for Steel Surface Defect Detection Algorithm Based on Context-Aware

This study addresses the challenges of steel surface defect detection, such as complex backgrounds and diverse defect morphologies, by proposing a context-aware YOLOv8-based algorithm. The algorithm introduces three key innovations: (1) Switchable Atrous Convolution (SAC) to enhance the receptive field for handling complex defects, (2) Attention Scale Fusion (ASF) with Context Anchor Attention (CAA) and Iterative Attentional Feature Fusion (iAFF) for improved multi-scale feature fusion and contextual dependency capture, and (3) SAC-based detection head for adaptive context awareness in classification and localization. Validation on the NEU-DET dataset shows a 6 % improvement in mAP and a 45.6 % reduction in computational load compared to the baseline, while maintaining real-time performance at 91 FPS. Generalization tests on the PCB dataset confirm the model's robust adaptability.