Answers are grounded in this paper and cite their sources.
Visual estimation of carbon fiber reinforced polymer (CFRP) countersunk-hole profiles is difficult when woven texture, weak contrast and contamination obscure circular boundaries. We develop a framework combining continuous edge confidence with a shared-center model. A SimpleUNet with 1.93 million trainable parameters predicts soft-edge responses. Confidence-ranked progressive sampling initializes the outer circle, a concentric histogram initializes the inner radius, and count-balanced confidence-weighted least squares refines one center and two radii. The dataset comprises 765 physical holes from 85 plates, partitioned at plate level into mutually exclusive training, validation and test sets in a 6:2:2 ratio. Model and parameter selection use the training and validation partitions. Joint success requires valid dual profiles and all four center and radius errors relative to separately annotated visual references to be strictly below 5.0 pixels. Held-out test success is 78.43% overall, including 88.89% for clean and 66.67% for contaminated surfaces. Controlled refinement comparisons distinguish balancing from confidence weighting. Paired errors and threshold responses reveal parameter-specific trade-offs. The results support local profile estimation under the evaluated conditions; independent dimensional validation, uncertainty propagation and comparisons with direct learning-based circle or ellipse detectors remain necessary for broader assessment.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: