Convolutional Neural Networks for Line Detection in Industrial Images

This thesis examines automatic line detection in industrial images acquired by optical measuring systems, with the aim of supporting the inspection and measurement of mechanical components. The proposed approach adapts a convolutional neural network in which candidate lines are represented in a discretized Hough Space, classified through confidence scores, and refined by regression. As part of the internship, a dedicated dataset was constructed from industrial images acquired at Vici \& C. S.p.A. using the MTL systems and annotated with a custom tool developed for this project. Finally, the thesis investigates a topology-based extension of the model by acting directly on the training loss, introducing additional regularization terms to encourage more coherent predicted line configurations. The experimental results show that the revised model detects relevant profile lines, while topology-aware extensions provide additional control over the structure of its predictions, representing a promising basis for further development. However, accuracy, robustness, and generalization remain aspects that need to be improved, particularly in more challenging images with a high density of lines.