A Steel Strip Defect Detection Method Integrating Spatial Reconstruction and Adaptive Feature Enhancement

Abstract

Steel strip surface defect detection plays a crucial role in quality control in industrial manufacturing. However, existing methods often face difficulties in achieving a balance between detection accuracy and efficiency in complex defect scenarios, particularly for multi-scale defects and intricate textured backgrounds. To address these challenges, this paper proposes an improved YOLOv11-based model, termed SAB-YOLO, which enhances detection performance by improving feature extraction and multi-scale feature fusion. Specifically, in the feature extraction stage, a Spatial Reconstruction Convolution (SCRConv) module is designed to mitigate the loss of fine-grained information during downsampling. In the feature fusion stage, a weighted fusion structure based on BiFPN is employed to improve multi-scale feature interaction. In the detection stage, a decoupled detection head, ASFFHeadV2, integrating the ASFF mechanism, is developed to enhance the model’s ability to detect small-scale and densely distributed defects. Extensive experiments on the NEU-DET and GC10-DET datasets demonstrate that SAB-YOLO achieves significant improvements over YOLOv11n in mAP@0.5 and recall while maintaining comparable inference efficiency. Moreover, it exhibits strong robustness in cross-dataset evaluations. The proposed method shows excellent adaptability in complex industrial defect detection tasks, providing a high-precision solution for automated steel strip surface inspection.