Improved YOLOv11based on multi-receptive field feature extraction and lightweight cross attention fusion for strip steel surface defect detection

An improved YOLOv11 (MRFFELCA-YOLO) algorithm based on Multi receptive Field Feature Extraction (MRFFE) and Lightweight Cross Attention Module (LCA) was proposed in order to address the problem of difficult detection of target defects caused by complex background interference such as varying sizes, diverse types, and insufficient lighting on the surface of strip steel. Firstly, in response to the insufficient adaptability of YOLOv11 feature extraction stage to multi-scale defects, an MRFFE module was designed to enhance the model's feature recognition ability for defects of different scales through multi branch parallel convolution combined with residual flow. Secondly, in order to address the issues of information dilution and noise introduction caused by simple concatenation during the YOLOv11 feature fusion stage, an LCA module was introduced to construct cross attention and feature enhancement layers to achieve adaptive feature interaction and suppress background interference, thereby improving the recognition accuracy of defect features. Thirdly, to address the issue of sensitivity to small targets and low localization accuracy in Intersection over Union (IoU) based loss functions, Gaussian Combination Distance (GCD) with scale invariance and joint optimization mechanism is adopted as the loss function to enhance the algorithm's ability to detect defects in small targets. Finally, testing was conducted on NEU-DET (Northeastern University Strip Defect Dataset), and the results showed that the mean Average Precision (P_mAP) of the proposed algorithm achieves 78.61%, representing a 3.71% improvement over the baseline YOLOv11.