An image segmentation method for metal parts under complex circumstances based on deep learning

Machine vision is often used to measure metal parts in three dimensions. However, accurately recognizing laser stripes that illuminate the surface of the part can be challenging due to the curvature and special materials of the metal parts that produce reflections, interference spots, strong ambient light, and other effects. Precisely segmenting the target area in the presence of these challenges is a crucial requirement. To address this problem, an improved U-Net semantic segmentation algorithm has been proposed in this paper for accurately segmenting laser stripes. The algorithm has been tested on laser images of shafts and blade parts, and the experimental results demonstrate that it can obtain more complete, smoother, and denser image segmentation results than traditional methods even in instances of highly reflective surfaces, significant interfering spots, and strong ambient light. These results verify the feasibility of the proposed method.