Abstract: Detecting surface defects in cold-rolled steel strips efficiently is crucial for quality control. However, such defects often exhibit multi-scale characteristics and are frequently obscured by complex backgrounds. To address this challenge, this paper proposes a lightweight detection framework—EAF-YOLO. The method incorporates an Efficient Spatial Coordinate Attention Module (ESCAM) to enhance the model's ability to perceive defect regions under complex backgrounds; a Multi-scale Feature Extraction module (MFE) to improve the representation of defects at different scales; and an Adaptive Attention Convolution module (AAConv) to enable dynamic fusion and enhancement of multi-scale features. Experimental findings demonstrate that EAF-YOLO maintains low computational complexity while attaining an average accuracy (mAP50) of 88.0%, striking a balance between high accuracy and minimal computational overhead. Results from on-site deployment thoroughly validate the practicality of the EAF-YOLO model in real production settings, with classification accuracy for various defects exceeding 90%.

Key words: surface defect detection, convolutional neural network, multi-scale feature extraction, attention mechanism