计算机技术与发展2025，Vol.35Issue(11)：46-52,7.DOI:10.20165/j.cnki.ISSN1673-629X.2025.0155
基于掩码区域一致性的半监督语义分割
Semi-supervised Semantic Segmentation Based on Mask Region Consistency
摘要
Abstract
The existing semi-supervised semantic segmentation methods mainly rely on single pixel or global features for supervision,but have limited utilization of the spatial structure of local regions,resulting in certain limitations of the model in distinguishing appearance similar categories.To this end,we propose a semi-supervised semantic segmentation method based on mask region consistency(Mask-Match).By introducing a mask region consistency module,the spatial relationships in unlabeled data are used as additional supervised signals to enhance the model's perception of local structures.In addition,to further improve the performance of the module,an uncertainty driven resampling strategy is proposed,which prioritizes unlabeled samples with high prediction uncertainty during the training process to enhance the model's learning ability for difficult instances.Specifically,the proposed method randomly masks certain regions on unlabeled images and constrains the model's prediction results to be consistent with the pseudo labels of the complete image,enabling the model to infer the semantic category of the masked regions using contextual information,thereby improving the robustness of feature expression.The extensive experiments on two benchmark datasets,Pascal and Cityscapes,have showed that the proposed method can ef-fectively improve segmentation performance under multiple label ratio settings,verifying its effectiveness and superiority.关键词
半监督语义分割/空间上下文建模/掩码区域一致性/重采样/伪标签Key words
semi-supervised semantic segmentation/spatial context modeling/mask region consistency/resampling/pseudo labels分类
信息技术与安全科学引用本文复制引用
黄新桃,张鸿..基于掩码区域一致性的半监督语义分割[J].计算机技术与发展,2025,35(11):46-52,7.基金项目
国家重点研发计划(2020AAA0108503) （2020AAA0108503）
