A Metal Grain Boundary Extraction Algorithm Based on Improved U-Net and Grain Boundary Repair

The microstructural characteristics of metal materials and their distribution are closely related to their chemical composition, processing, and performance. Accurate and comprehensive extraction of grain boundaries is a prerequisite for extracting microstructural features, which requires precise segmentation of grains and grain boundaries. However, the complexity of grain boundaries in microstructure images of metal materials makes precise segmentation challenging. This paper proposes an improved U-net network, achieving better segmentation performance compared to existing methods. Nonetheless, defects such as fractured grain boundaries still exist in the results. To address these defects, we further developed a grain boundary repair algorithm that incorporates materials domain knowledge to ensure the extraction of accurate and comprehensive grain boundaries. Experimental results validate the effectiveness of the proposed algorithm.