Research on Interactive Segmentation-Based Mask Image Generation Technology in the Domain of Precision Assembly

Computer vision has become a crucial tool, demonstrating significant performance in precision assembly tasks. However, as manufacturing processes become increasingly complex, data-driven neural network models require substantial amounts of high-quality training data. Manual annotation process involves meticulously delineating marking points at the pixel level to outline polygonal objects, making it labor-intensive and costly. For intricate objects, this poses a challenge to the efficiency and quality of engineers’ annotated masks. This paper introduces a novel approach that leverages fine-tuning techniques applied to the Segment Anything Model (SAM), incorporating additional mask encoding information. Across various image tasks within the manufacturing pipeline, our method demonstrates a 3.0% improvement in accuracy compared to using SAM alone. The semi-automatic annotation scheme proposed in this paper aims to assist engineers in enhancing annotation efficiency.