Abstract

Steel manufacturers currently rely on optical sensing systems that capture surface images for quality control, but these systems still face practical challenges in industrial environments. Early machine vision methods used handcrafted features, and their accuracy dropped when lighting changed or the background texture became complex. Deep learning improved detection, yet small defects that blend into background noise or share visual patterns with other classes still cause missed detections and false positives on factory floors where hardware resources are tight. GEA-YOLO integrates different refinement strategies into the Backbone, Neck, and training stage. C2DGA replaces C2PSA in the Backbone. Deformable attention adapts to defect structures that deviate from fixed sampling patterns, and a dynamic gate fuses global contextual information with local texture features. EMA modules are inserted into the Neck, where they recalibrate features independently at each scale and reduce the influence of background interference. DetectAux provides auxiliary supervision for hard samples during training. Unlike approaches that introduce attention at a single fixed position, GEA-YOLO places each module at the stage where it can improve the corresponding representation. We evaluated GEA-YOLO on the NEU-DET and GC10-DET datasets. On NEU-DET, the model reached 80.3% mAP@0.5, 2.6 percentage points above YOLOv11s, while keeping a reasonable balance between accuracy and computational cost and maintaining an inference speed of 169.5 FPS. Cross-dataset validation on GC10-DET further confirmed generalization, yielding 74.9% mAP@0.5, 2.9 percentage points above YOLOv11s, showing strong potential for real-time steel surface inspection. Ablation results confirm that each modification fixes a different weakness in the detection pipeline, but the full gain only appears when all three are used together. These results indicate that GEA-YOLO is promising for real-time optical inspection in controlled benchmark settings.

IPC Classification

Keywords