关键词: 钢材表面缺陷检测, YOLOv11, 双流架构, 多模态

Abstract: To address complex background interference and defect diversity in steel surface defect detection,a steel surface defect detection method based on improved YOLOv11 and a multimodal fusion mechanism was proposed.Firstly,a dual-stream architecture network model was designed based on YOLOv11,which used to extract multi-modal feature information from RGB image stream and depth image stream,respectively. The dilated convolution was used to expand the receptive field of the depth image,and the depth enhancement module adaptively adjusted the weights of each channel to enhance the perception ability of the geometric structure of the steel surface. A self-attention mechanism was used to focus on critical defect areas and enhance the fusion of multimodal information.Next,the gradient aware region balance loss function was designed,which introduced gradient consistency loss and feature alignment loss,and weighted and summed multiple loss terms to improve the accuracy,robustness,and adaptability of steel surface defect detection. Finally,an adaptive multimodal collaborative quantization strategy was designed,which used to ensure improved inference speed of the dual-stream network model without sacrificing detection accuracy. The experimental results showed that the proposed improved method had better detection performance,compared to several mainstream defect detection algorithms,and could balance detection accuracy and speed well. The mAP reached 92.1 %,and the detection speed reached 48.3 FPS,showing improvements of 6.7 % and 11.8 FPS over the original YOLOv11,respectively. The visualization results demonstrated it′s adaptability to complex environments and diverse defect types,supporting automated quality inspection in steel manufacturing.

Key words: steel surface defect detection, YOLOv11, dual-stream architecture, multimodal