A joint image enhancement method using A-MSR and AG-ABF for regionally heterogeneous steel plate surface defects

Traditional enhancement methods may amplify rolling textures and background fluctuations when weak defects coexist with regionally heterogeneous steel-surface patterns. To address this problem, an adaptive dual-domain fusion framework (ADF) is proposed for hot-rolled steel surface-defect images. A variance-driven adaptive multi-scale Retinex module (A-MSR) maps the local variance of the normalized MSR response to a bounded gain field and regulates the local detail residual while preserving the local mean. A gradient-guided adaptive bilateral-filtering module (AG-ABF) combines small- and large-scale bilateral responses according to local edge strength, providing an auxiliary edge-preserving smoothing response. The two branches are decomposed into structure and detail layers and integrated using spatially varying weights determined by relative detail variance and edge guidance. On NEU-DET, ADF achieved an average gradient of 0.5715, an information entropy of 7.1948, and a pooled edge-to-background ratio of 1.3806. Compared with A-MSR, ADF reduced the background-region gradient by 20.5% while retaining approximately 80.8% of the defect-region gradient and 97.5% of the information entropy. In downstream YOLOv8n detection experiments, ADF improved mAP50 and mAP50–95 by 1.73 and 2.08 percentage points, respectively, compared with Raw. On GC10-DET, ADF achieved an mAP50 of 0.7277 and an mAP50–95 of 0.3545, outperforming Raw and AW-MSR in both mAP metrics. These results indicate that ADF provides a favorable trade-off among defect enhancement, background suppression, information preservation, and downstream detection performance.