YOLOv11-WBD: A wavelet-bidirectional network with dilated perception for robust metal surface defect detection

Abstract

In the field of quality control, metal surface defect detection is an important yet challenging task. Although YOLO models perform well in most object detection scenarios, metal surface images under operational conditions often exhibit coexisting high-frequency noise components and spectral aliasing background textures, and defect targets typically exhibit characteristics such as small scale, weak contrast, and multi-class coexistence, posing challenges for automatic defect detection systems. To address this, we introduce concepts including wavelet decomposition, cross-attention, and U-shaped dilated convolution into the YOLO framework, proposing the YOLOv11-WBD model to enhance feature representation capability and semantic mining effectiveness. To improve robustness, a plug-and-play Wavelet-Attentive Multiband Fusion Module (WAMF) is designed, achieving decoupling of low-frequency and high-frequency features and adaptive multi-frequency fusion. To effectively aggregate multi-scale features, a Bottleneck-Enhanced Dilated U-Conv Module (BEDU) is designed, fusing global and local information with lower computational resource consumption. To address feature fusion, a Bidirectional Depthwise Cross-Attention Module (BDCA) is designed to replace simple concatenation and convolution operations, achieving adaptive feature fusion. YOLOv11-WBD undergoes rigorous evaluation on the public NEU-DET and GC10-DET datasets; experimental results show that the improved model achieves performance gains on both datasets: the mAP@0.5 metric increased by 5.8% and 2.8% respectively. Furthermore, the improved model demonstrates stronger noise tolerance, maintaining high defect detection capability even in moderate noise environments, providing a valuable solution for industrial applications.