RLL-YOLO-seg: An Improved Instance Segmentation Model for Pixel-Level Steel Surface Defect Detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Steel surface defect detection is of great significance for ensuring the quality of industrial products and the safety of production. To address the issue that traditional steel surface defect detection methods often rely on object localization and lack finegrained description of defect regions, this paper proposes an improved instance segmentation model, RLL-YOLO-seg, which enhances the detection accuracy of steel surface defects through pixel-level segmentation. First, the C3K2-RVB lightweight feature extraction module, through multi-scale convolution, residual structure, and Squeeze-and-Excitation, improves the perception of tiny defects and local textures.Second, the SPPF-LSKA attention enhancement module, combining spatial pyramid pooling and large separable kernel attention, optimizes multi-scale feature fusion and the capture of defect edge details. Finally, the Segment-LQE quality-aware segmentation head, by decoupling object localization from mask generation, improves segmentation accuracy and boundary integrity. Experimental results show that RLL-YOLO-seg achieves a Box mAP of 83.6% and a Mask mAP of 76.9% on the NEU-Seg dataset, which are 5.2% and 4.1% higher than those of the baseline model, outperforming several mainstream instance segmentation methods. Meanwhile, on the GC7 dataset, RLL-YOLO-seg achieves a Box mAP of 84.0% and a Mask mAP of 80.0%, representing improvements of 0.7% and 3.8% over the baseline model, demonstrating good generalization ability and robustness. In summary, RLL-YOLOseg can accurately localize and segment steel surface defects, providing strong support for quantitative analysis and intelligent quality assessment of steel surface defects.