GSLA-YOLOv11: Lightweight Steel Surface Defect Detector with Gated Multi-Branch Aggregation

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).

Reliable steel surface defect detection remains challenging because defects often exhibit weak texture contrast, large scale variation, elongated morphology, and complex background interference. To address these problems, this study proposes GSLA-YOLOv11, a lightweight detector based on YOLOv11s. In the backbone, a gated multi-branch aggregation module is introduced to enhance local texture, contextual, and directional feature representation. In the neck, a scale-aware fusion module is designed to improve aligned cross-scale feature interaction. Before the detection head, a lightweight channel attention enhancement module recalibrates informative channels to support more stable defect classification and localization. Experiments on the NEU-DET dataset show that GSLA-YOLOv11 achieves 88.6% mAP@0.5 and 55.3% mAP@0.5:0.95, improving mAP@0.5 by 8.3 percentage points over YOLOv11s while using only 1.8 × 10⁶ parameters and 8.8 × 10⁹ FLOPs. On the GC10-DET dataset, the proposed model achieves 75.5% mAP@0.5, outperforming YOLOv11s by 2.0 percentage points. Deployment experiments further show that GSLA-YOLOv11 reaches 161.3 FPS under PyTorch FP32 inference and 333.3 FPS under TensorRT FP16 inference on an RTX 4060 Laptop GPU. On a Jetson Orin NX 16 GB, it achieves 116.3 FPS with 780 MB peak memory usage and 20.5 W average power consumption under TensorRT FP16 inference. These results demonstrate that GSLA-YOLOv11 provides an effective trade-off among detection accuracy, model compactness, and real-time deployment capability for steel surface defect inspection.