Refining UNet3+ for Segmentation of Microstructures in Steel Images

Segmentation of microstructures in steel images is crucial as it helps in assessing the quality and strength of the material. Traditional and existing deep learning models often struggle to capture both the textural details and structural patterns of the microstructures found in steel images. In this paper, we enhance the structural representation of steel images by modifying the UNet3+ design, including data processing, architectural modules, and a tailored loss function. Experimental results show that the modified model is effective in distinguishing microstructures not only within the same steel type and magnification, but also across different scales and steel types.