Boundaries Matters: A Novel Multibranch Semisupervised Semantic Segmentation Method

In recent years, semisupervised semantic segmentation (SSS) research has been progressing rapidly. Existing methods usually ignore the classification of detailed pixels, such as boundaries, resulting in degraded segmentation performance. To overcome this challenge, we propose a new multibranch SSS framework, BoundaryMatch, which combines image- and feature-level perturbation branches with a boundary detail guidance branch, all utilizing a shared encoder. Specifically, this boundary module enhances segmentation by integrating the learning of spatial information into low-level layers in a single-stream manner. Finally, the low-level features and deeper features are fused together to predict the final segmentation result and achieve accurate correction of boundary pixels. This multibranch approach improves on the shortcomings of consistency regularization that focus only on maintaining the global consistency of the image. Extensive experiments conducted on the Cityscapes and PASCAL VOC 2012 datasets demonstrate that the method proposed in this article effectively enhances the performance of SSS.