Signal, Image and Video Processing Обработка сигналов
MTF-YOLO: a steel surface defect detector with multiscale texture fusion and variance-guided convolution
Аннотация
In industrial surface inspection, defects are often small, weakly textured, and bounded by uncertain edges—conditions under which state-of-the-art detectors struggle, especially in low-contrast scenes and for thin cracks. To address these challenges, this study proposes MTF-YOLO, a YOLO-based detector that integrates three modules—Defect Variance-Guided Convolution (DVGC), Defect Edge Enhancement (DEE), and Defect-Oriented Texture Fusion (DOTF). First, DVGC explicitly models the statistical disparity between foreground and background, improving separability for small, low-contrast targets. Second, DEE sharpens contours and fine details at the feature level, mitigating localization drift caused by blurred boundaries. Third, DOTF aggregates texture cues across multiple scales and orientations. In addition, the proposed architecture employs a Path Aggregation Network–FPN (PAN-FPN) neck together with Spatial Pyramid Pooling-Fast (SPPF) to fuse semantic and low-level information efficiently across pyramid levels. Experiments show that MTF-YOLO achieves mAP@0.5 scores of 83.6±0.4% and 82.3±0.3% on NEU-DET and GC10-DET, respectively (mean±std over three random seeds); it strikes a favorable balance among accuracy, efficiency, and deployability, and demonstrates robust performance for steel surface defect detection.