MZMingxiang Zhu
Mingxiang Zhu
*This is an algorithmically generated author record based on publicly available information.
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Mingxiang Zhu has a total of 4 co-authors. Here are the top co-authors by number of shared publications:
- Hongyan Sun
School of Electric Power Engineering, Nanjing Normal University Taizhou College, Taizhou 225300, China
1 shared publication - Jue Wang
School of Electric Power Engineering, Nanjing Normal University Taizhou College, Taizhou 225300, China
1 shared publication - Yu Wang
School of Electric Power Engineering, Nanjing Normal University Taizhou College, Taizhou 225300, China
1 shared publication - Zhizhen Li
School of Electric Power Engineering, Nanjing Normal University Taizhou College, Taizhou 225300, China
1 shared publication
Latest Publications
- Open Access
TDRPFormer: Texture-Debiased Reliable Prototype Transformer for Pixel-Level Steel Surface Defect Segmentation
Steel surface defect segmentation is a critical task in industrial visual inspection, with direct implications for quality assessment, process control, and in-service safety. However, complex steel-surface images often contain strong background textures, local reflections, machining traces, and low-contrast defects, which cause existing methods to be distracted by normal textures and to produce false positives or missed detections in slender structures, blurred boundaries, and weak-response regions. To address these challenges, we propose a Texture-Debiased Reliable Prototype Transformer (TDRPFormer) for pixel-level steel surface defect segmentation. Within a query-driven mask prediction framework, TDRPFormer first derives pixel-level confidence and boundary cues from intermediate segmentation priors and then aggregates normal texture prototypes from high-confidence background regions. A Texture Debiasing Module (TDM) is introduced to suppress the interference of normal textures in defect representations. Next, a Reliable Prototype Routing (RPR) module compresses the debiased spatial features into a small set of high-reliability prototypes, thereby reducing ineffective interactions between queries and redundant background pixels. Finally, a Structural Continuity Modulation (SCM) module enhances high-resolution mask features, improving the recovery of slender defects, weak-response regions, and blurred boundaries. We conduct systematic experiments on the ESDIs-SOD and NEU-DET steel surface defect datasets. TDRPFormer achieves the best overall performance on both datasets. On ESDIs-SOD, the mean absolute error (MAE), weighted F-measure (), S-measure (), and mean E-measure () reach 0.0185, 0.8824, 0.9070, and 0.9626, respectively. On NEU-DET, the corresponding values are 0.0210, 0.8858, 0.9017, and 0.9662. Further ablation studies and response visualizations confirm the complementary roles of texture debiasing, reliable prototype routing, and structural continuity modulation. These results demonstrate that TDRPFormer improves discriminability, stability, and structural recovery under complex industrial backgrounds, providing an effective solution for pixel-level steel surface defect segmentation.Open AccessTDRPFormer: Texture-Debiased Reliable Prototype Transformer for Pixel-Level Steel Surface Defect Segmentation
Steel surface defect segmentation is a critical task in industrial visual inspection, with direct implications for quality assessment, process control, and in-service safety. However, complex steel-surface images often contain strong background textures, local reflections, machining traces, and low-contrast defects, which cause existing methods to be distracted by normal textures and to produce false positives or missed detections in slender structures, blurred boundaries, and weak-response regions. To address these challenges, we propose a Texture-Debiased Reliable Prototype Transformer (TDRPFormer) for pixel-level steel surface defect segmentation. Within a query-driven mask prediction framework, TDRPFormer first derives pixel-level confidence and boundary cues from intermediate segmentation priors and then aggregates normal texture prototypes from high-confidence background regions. A Texture Debiasing Module (TDM) is introduced to suppress the interference of normal textures in defect representations. Next, a Reliable Prototype Routing (RPR) module compresses the debiased spatial features into a small set of high-reliability prototypes, thereby reducing ineffective interactions between queries and redundant background pixels. Finally, a Structural Continuity Modulation (SCM) module enhances high-resolution mask features, improving the recovery of slender defects, weak-response regions, and blurred boundaries. We conduct systematic experiments on the ESDIs-SOD and NEU-DET steel surface defect datasets. TDRPFormer achieves the best overall performance on both datasets. On ESDIs-SOD, the mean absolute error (MAE), weighted F-measure (), S-measure (), and mean E-measure () reach 0.0185, 0.8824, 0.9070, and 0.9626, respectively. On NEU-DET, the corresponding values are 0.0210, 0.8858, 0.9017, and 0.9662. Further ablation studies and response visualizations confirm the complementary roles of texture debiasing, reliable prototype routing, and structural continuity modulation. These results demonstrate that TDRPFormer improves discriminability, stability, and structural recovery under complex industrial backgrounds, providing an effective solution for pixel-level steel surface defect segmentation.