BCD-YOLOv11: an enhanced YOLOv11 model for steel surface defect detection
TL;DR
An enhanced BCD-YOLOv11 detection algorithm based on YOLOv11 is proposed, which achieves a 7.7% increase in precision and a 2.1% improvement in mAP while maintaining inference speed and shows particularly notable performance in detecting small and overlapping defects.
Abstract
To address the challenges of insufficient feature extraction, limited multi-scale defect detection capability, and poor localization accuracy in steel surface defect detection tasks, this paper proposes an enhanced BCD-YOLOv11 detection algorithm based on YOLOv11. The algorithm incorporates three key improvements: First, the CARAFE upsampling operator is introduced in the feature fusion stage, replacing traditional interpolation methods with a content-aware feature reorganization mechanism. This significantly enhances the upsampling quality of feature maps and improves the retention of fine defect features. Second, the BiFPN network is adopted to reconstruct the multi-scale feature fusion path. Through weighted bidirectional cross-scale connections, it enables efficient integration of deep and shallow features, thereby effectively enhancing the model's detection performance for defects of varying sizes. Finally, the Wasserstein Distance Loss is employed as the bounding box regression loss function. By directly minimizing the distribution distance between predicted and ground truth boxes, it improves the localization accuracy of defects in complex backgrounds. Compared to the baseline YOLOv11 model, BCD-YOLOv11 achieves a 7.7% increase in precision and a 2.1% improvement in mAP while maintaining inference speed. It shows particularly notable performance in detecting small and overlapping defects, demonstrating the effectiveness of the proposed enhancements.