Technical Summary: Morphology-Aware Sample Assignment for Surface Defect Detection

1. Problem Statement

The paper addresses a fundamental limitation in visual object detection, specifically within the context of industrial surface defect detection: the insensitivity of the Intersection-over-Union (IoU) metric during positive sample assignment.

While IoU is the standard criterion for evaluating the geometric alignment between candidate proposals and ground-truth annotations, theoretical modeling reveals the existence of a "non-sensitive region" (or plateau) on the IoU response curve. Within this region, candidate boxes with distinct geometric overlaps (e.g., different shapes, aspect ratios, or areas) yield nearly identical IoU scores. This phenomenon leads to:

• Assignment Ambiguity: The inability to distinguish subtle geometric differences among densely distributed candidates.
• Weak Gradient Propagation: Ambiguous sample ranking hinders effective optimization during training.
• Cross-Instance Errors: In scenarios with spatially close or overlapping ground-truth instances, IoU-based metrics may incorrectly assign samples to the wrong instances.

This limitation is particularly detrimental to industrial defect detection, where defects exhibit high diversity in scale and morphology, requiring fine-grained geometric alignment that standard IoU fails to provide.

2. Methodology: Morphological Characteristic Cost (MCC)

To overcome IoU's insensitivity, the authors propose a Morphology-Aware Sample Assignment framework centered on the Morphological Characteristic Cost (MCC). The methodology involves three core components:

A. Parameterized Candidate Box Modeling

The authors parameterize the candidate box space using $(x, y, \text{wh\_ratio}, \text{area})$. Through theoretical analysis and visualization, they demonstrate that for specific branches of this space (e.g., constant aspect ratio or constant area), the IoU curve exhibits flat plateaus where geometric variations do not alter the score.

B. Multi-Dimensional Similarity Metrics

To resolve the flat regions, the authors introduce three complementary morphological similarity metrics, each normalized to a range of $(0, 1]$:

1. Area Similarity ($Score_{Area}$): Measures the scale difference between the candidate and ground-truth areas.
   $$Score_{Area} = 1 - \frac{|Area_{candi} - Area_{gt}|}{\max(Area_{candi}, Area_{gt})}$$
2. Shape Similarity ($Score_{Shape}$): Evaluates the "slimness" or "squatness" by comparing the ratio of the shorter side to the longer side ($min(w,h)/max(w,h)$).
3. Width-Height Ratio Similarity ($Score_{WH}$): Directly compares the aspect ratios ($w/h$) to resolve the symmetry ambiguity inherent in the shape metric (where a square and a rectangle with swapped dimensions might yield similar shape scores).

C. The MCC Score and Final Similarity

The final morphological score is computed as the mean of these three metrics:
$$MCC_{Score} = \frac{1}{3} (Score_{Area} + Score_{Shape} + Score_{WH})$$

The overall similarity used for sample assignment is the product of the original IoU and the MCC score:
$$Similarity = IoU \times MCC_{Score}$$

Theoretical Impact:

• Gradient Enhancement: The combination of metrics reshapes the response distribution, eliminating flat plateaus and creating smooth, directional gradients.
• Polygonal Iso-Contours: Unlike the rectangular boundaries of IoU, the new similarity metric forms polygonal iso-response contours. This tightly confines high-response regions around the ground-truth instance, reducing the "blind zone" where ambiguous samples exist.
• Zero Overhead: The method is a "plug-and-play" modification to the sample assigner, requiring no changes to the model backbone, head, or inference pipeline.

3. Key Contributions

The paper outlines four primary contributions:

1. Theoretical Analysis: Construction of a parameterized framework that visualizes and proves the existence of IoU's insensitivity region, identifying it as a cause for assignment errors and degraded training efficiency.
2. MCC Mechanism: Proposal of a multi-dimensional similarity measure (Area, Shape, Aspect Ratio) that effectively compensates for IoU's limitations, yielding directional gradients and eliminating blind zones in candidate evaluation.
3. Improved Sample Selection: Demonstration that the refined matching landscape compresses high-response regions into compact, localized neighborhoods around ground-truth boxes, thereby enhancing the precision of positive/negative sample separation and reducing false positives.
4. Plug-and-Play Efficiency: Validation that the method integrates seamlessly into existing frameworks (specifically YOLOv9) with zero additional computational cost or inference overhead, making it highly suitable for industrial deployment.

4. Experimental Results

The authors evaluated the method using the YOLOv9 family of models on two industrial surface defect datasets: GC10-DET and NEU-DET.

• Performance Gains:
  • GC10-DET: The method achieved consistent mAP improvements across all YOLOv9 variants (t, s, m, c, e), with an average net gain of 2.38%. Notably, the YOLOv9-e model reached 74.2% mAP (a +2.5% gain).
  • NEU-DET: Similar stable improvements were observed, with an average mAP gain of 1.08% across the detector family.
• Training Efficiency: Training curves showed that the proposed method accelerated convergence. The mAP_0.5 threshold of 0.6 was reached by epoch 50 with MCC, compared to epoch 80 for the baseline.
• Defect-Specific Improvements: The method significantly improved detection for challenging, morphologically diverse defects (e.g., creases and rolled pits), as evidenced by Precision-Recall curve analysis and feature heatmap visualizations showing higher activation alignment with ground truths.
• Comparison with SOTA: When compared against other advanced defect detectors (e.g., GSCNet, LiFSO-Net, RT-DETR), the proposed method achieved competitive or superior absolute mAP scores while maintaining the original model architecture. The authors argue that achieving high performance on a simpler baseline (FCOS/YOLOv9) without structural modifications demonstrates the method's advancement and generality.

5. Significance and Claims

The paper claims that the proposed approach offers a data-efficient training paradigm that addresses a theoretical flaw in standard detection metrics without compromising deployment speed.

• Industrial Relevance: The method is specifically tailored for industrial visual inspection where defects vary drastically in scale and shape. The ability to distinguish subtle geometric differences is critical for high-precision localization.
• Deployment Advantage: By requiring zero additional inference overhead and no structural modifications, the method is presented as a practical solution for real-world systems where computational resources and inference latency are constrained.
• Generality: The authors assert that the method is not limited to YOLOv9 but is applicable to any detection framework utilizing sample assignment (e.g., RetinaNet, ATSS), as it operates solely at the similarity calculation level of the assigner.

The work concludes that optimizing the positive sample selection strategy through morphological awareness is a viable and effective path to enhancing detection performance for multi-scale and morphologically diverse targets.