Domain adaptive segmentation method for mechanical assembly based on iterative loops

During the assembly process of mechanical products, employing deep learning techniques for the semantic segmentation of assembly images enables real-time monitoring of irregularities, including incorrect or missing assemblies. However, most of the current monitoring methods based on deep learning adopt supervised learning. This requires a large number of labels according to different assembly specifications, which is time-consuming and laborious. To address this issue, this study designed a two-stage adaptive segmentation framework based on iterative loops for synthesis-physical assembly images, i.e., ILDA-Net (iterative loops domain adaptation network), which does not require any labeling of physical assemblies. In the adversarial learning stage, a trainable line-guided filter module and a line discriminator module are introduced for maintaining line features. The two modules are iteratively trained in a loop to continuously optimize the segmentation model. In the self-training stage, the edge segmentation quality is guaranteed by optimizing the segmentation model through utilizing unreliable pseudo-labels. Finally, this study constructed a set of semantic segmentation datasets for domain adaptation of synthetic-physical assembly images and conducted experiments on these datasets. Based on these experiments, the Dice coefficient can reach up to 89.33%, which demonstrating that the proposed method can be utilized for the physical assembly image segmentation.