Morphology-aware assignment boosts defect detection with zero overhead

A new paper from researchers Pengfei Liu and Yuhan Guo tackles a fundamental flaw in object detection for industrial surface defect detection: IoU insensitivity. Intersection-over-Union (IoU) is the standard metric for matching candidate proposals to ground-truth annotations, but the authors mathematically model a 'non-sensitive region' where samples with significantly different geometric overlaps produce nearly identical IoU scores. This leads to poor positive sample assignment and reduced training efficacy.

To fix this, the team introduces a set of morphological similarity metrics that capture area, shape, and aspect ratio. These are aggregated via mean-based scoring into a supplementary matching score that compensates for IoU’s blind spots. Theoretically, this reshapes the response distribution to create polygon-like iso-response contours that tightly confine high-response regions around each ground-truth instance. Empirically, integrating the approach into YOLOv9 delivers consistent performance gains on the NEUDET and GC10-DET datasets. Crucially, the method is fully plug-and-play with zero additional inference overhead, making it directly deployable in real-time industrial visual inspection systems.