ECD-YOLO: Multiscale Feature Analysis and Detail-Preserving Downsampling for Steel Surface Defect Detection

Abstract

: Automated inspection of steel surfaces is challenged by minute defects, weak contrast, repetitive rolling textures, and substantial scale variation. This paper presents ECD-YOLO, an enhanced YOLO11n detector for accurate and efficient defect localization. The proposed ETC-Down module decomposes downsampling into four complementary branches that preserve spatial structure, gradient edges, multiscale texture, and local contrast. The DAFB module employs differential attention to suppress redundant background responses and combines large-kernel positional encoding, dilated depthwise convolution, and channel gating to strengthen global-local discrimination. SSFF and TFE are incorporated into the neck to align and fuse P3-P5 features across spatial scales. Experiments on the NEU-DET dataset show that, relative to YOLO11n, ECD-YOLO improves precision, recall, mAP@0.5, and mAP@0.5:0.95 by 12.1, 2.6, 5.4, and 10.2 percentage points, respectively, while increasing the parameter count from 2.58 M to 3.28 M and computation from 6.3 to 7.4 GFLOPs. Ablation results confirm complementary gains in detail preservation, background suppression, and multiscale fusion. ECD-YOLO therefore provides a favorable accuracy-efficiency trade-off for online steel surface inspection.