Semi-supervised Metallographic Image Segmentation via Consistency Regularization and Contrastive Learning

Metallographic image segmentation is a core task towards the automation of metallographic analysis. Currently, most advanced methods for this generally employ supervised deep learning models that require great number pixel-level annotated images, while annotation process time-consuming and labor-intensive. In order to address issue, semi-supervised model called Con2Net proposed in work, which leverages unlabeled data improve performance segmentation. The adopts multi-decoder architecture, enforces consistency constraint between each decoder’s output other decoders’ soft pseudo labels produced by sharpening. addition, mitigate negative impact caused sharpening on false predicted pixels, we adopt operation only accurately pixels. For labeled simplest effective way select pixels directly comparing them with ground-truth labels. contrastive module introduced, encourages have better intra-class compactness inter-class dispersion feature space. Based that, pseudo-labels are obtained calculating maximum similarity vectors, then could be filtered out. We conduct experiments two public datasets segmentation, five state-of-the-art three partition protocols. results demonstrate not outperforms baseline significant margin, but also achieves superior compared models. Our source code available at https://github.com/Siiimon2423/Con2Net.