A Comparative Study of Conventional and Deep Learning Based Computer Vision Methods for Circular Bore Detection in High-Variance Low-Volume Assembly Automation

This preprint compares five computer vision methods for detecting circular bores on four heterogeneous industrial panels, in the context of robotic pick-and-place automation for high-variance low-volume assembly. Three classical approaches: Circular Hough Transform, Gabor filter banks, and gradient-based edge detection are benchmarked against two single-stage deep learning detectors, SSD and YOLOv2, both implemented with a ResNet-50 backbone via transfer learning. All methods are trained and evaluated on the same dataset of 653 images using a unified protocol (precision, recall, F1, IoU, FPS). YOLOv2 with ResNet-50 (75 epochs) achieves the best trade-off between F1 (0.972) and throughput (≈ 11.6 FPS); SSD with ResNet-101 yields the most accurate bounding boxes; among classical methods, the gradient pipeline offers the best F1-vs-speed balance. A single neural detector trained on all four panels generalises without panel-specific tuning, whereas classical pipelines require panel specific pre-processing. The work was originally conducted in 2021 as part of a Master's thesis at the University of Luxembourg and is released as a preprint in 2026.