Morphology-Aware Sample Assignment: Overcoming IoU Insensitivity for Surface Defect Detection
A new research paper introduces Morphology-Aware Sample Assignment (MASA) to improve surface defect detection in visual models. MASA addresses the limitations of the Intersection-over-Union (IoU) metric by incorporating morphological similarity metrics for area, shape, and aspect ratio. This refinement ensures more accurate matching of candidate proposals to ground-truth annotations, leading to better training efficacy. Experiments using the YOLOv9 framework on NEUDET and GC10-DET datasets show performance gains without increasing inference overhead. AI
IMPACT Improves accuracy in AI-powered visual inspection for industrial defect detection.