Vision Transformers for Line Segment and Circular Arc Detection in Industrial Environments

This thesis focuses on the detection of geometric primitives in industrial images acquired by optical measuring machines. The objective is to detect line segments, circular arcs, and circles, which are essential to describe the profiles of industrial components. The main contributions of the work are the following. First, two dedicated industrial datasets were constructed and annotated: one for line segment detection and one for circular arc and circle detection. Second, the Deformable Transformer-based Line Segment Detector was adapted to the industrial line segment detection setting, and a revised formulation was developed to detect line segments, circular arcs, and circles within a unified geometric representation. The experimental results show that the proposed approaches can detect relevant geometric structures in industrial images and represent a promising basis for further development. Several directions remain open for increasing the accuracy, robustness, and generalization capability of the models.