New Method Enhances Defect Detection by Improving IoU Sensitivity

A new research paper highlights a significant limitation in current visual detection models, specifically concerning the Intersection-over-Union (IoU) metric. IoU, crucial for evaluating spatial alignment between proposals and ground-truth annotations, suffers from a "non-sensitive region" where distinct geometric overlaps yield nearly identical IoU scores. This insensitivity can degrade the quality of positive sample sets and hinder training efficacy. To overcome this, the researchers propose a morphology-aware sample assignment method. This approach introduces a set of morphological similarity metrics, encompassing area, shape, and aspect ratio, to enhance the positive sample selection process. A supplementary matching score, derived from the aggregation of these multidimensional similarities, compensates for IoU's inherent inability to fully represent structural correspondence. The theoretical underpinnings show that incorporating morphological similarity reshapes the matching function's response distribution, creating more effective directional gradients and polygon-like iso-response contours. These contours tightly confine high-response regions around ground-truth instances, substantially boosting the precision of positive sample selection. Demonstrated on the YOLOv9 framework, the method consistently improves performance on NEUDET and GC10-DET datasets, offering a plug-and-play solution with zero additional inference overhead, making it highly efficient for industrial visual inspection.