SCD-YOLO: Frequency-Domain Cleansing and Refinement for Lightweight Steel Surface Defect Detection

Abstract

Surface defect detection underpins automated quality control in modern manufacturing, where inspection must be both accurate and lightweight to keep pace with continuous production lines. Detecting flaws on continuous metal sheets is difficult because the captured images are dominated by intense periodic background textures and contain low-contrast defects with weak gradients, both of which confound conventional spatial-domain feature extractors. We propose SCD-YOLO, an efficient end-to-end detector built on the YOLO26n baseline. We replace the C2PSA attention block at the backbone tail with the Orthogonal Spectral Decoupling Module (OSDM), which suppresses dominant background frequencies in an amplitude-only, sample-adaptive manner while preserving phase. Two Fractional-Order-Inspired Convolution Refinement (FOCR) modules are inserted at the P4 and P5 branches of the neck, applying a bounded fractional-order-inspired amplification that sharpens the weak gradients of low-contrast defects. The two modules follow a cleanse-then-refine principle, in which OSDM first reduces the dominant background and FOCR then amplifies the surviving defect signal. On the GC10-DET test set, SCD-YOLO achieves the highest mAP@0.5 of 68.8% and mAP@0.5:0.95 of 35.4% among all compared methods, improving the YOLO26n baseline by 3.2 and 1.9 percentage points, respectively, while using only 2.24M parameters and 5.1 GFLOPs, both below the baseline’s 2.38M and 5.2 GFLOPs. On the NEU-DET test set, SCD-YOLO achieves 77.7% mAP@0.5 and the highest mAP@0.5:0.95 of 44.8%, improving the baseline by 0.9 and 0.6 percentage points, respectively, showing that the architectural benefit remains effective on a different steel surface defect distribution.