<p>Semi-supervised learning for image segmentation using pseudo-labels is promising. However, current methods typically use strict thresholds to select reliable pseudo-labeled pixels, leading to limited use of these labels. Relying only on highly reliable pseudo-labeled pixels can also hinder the model’s learning ability in challenging regions such as edges and minority categories. In order to explore other regions in unlabeled datasets, especially the transition region between objects and backgrounds, we propose an effective semi-supervised segmentation method called MaskMatch. Specifically, we introduce two methods to enhance the segmentation of the transition region between objects and backgrounds. First, we propose a dynamic background region mask. For labeled data, a potential error area mask (MPEA) is created, directing the model to focus on samples that are prone to segmentation error regions. For unlabeled data, an Uncertainty Correction Mask (MU) is introduced to explore the correct segmentation regions in regions below a specified threshold, thus improving the efficiency of pseudo-labeling. Second, the feature space gap between objects and backgrounds is increased by Object-Background Contrast Learning (OBCL), which improves the segmentation results in these transition regions. Extensive experiments on datasets such as PASCAL VOC 2012, COCO, LEVIR, and WHU have shown that MaskMatch delivers excellent results.</p>