Improving Semantic Segmentation with Semi- Supervised Learning Techniques

Consistency regularization has become an effective approach for improving semantic segmentation performance by encouraging stable predictions under different image transformations. However, many existing methods mainly concentrate on prediction-level consistency and optimize the complete network jointly, which may limit the effective use of supervisory information. To overcome this limitation, the proposed Multi-Constraint Consistency Learning (MCCL) framework enhances both the encoder and decoder stages through a feature consistency strategy that improves feature alignment and intra-class representation similarity. In addition, an adaptive noise intervention module is introduced to strengthen model robustness by enabling reliable predictions even when feature representations are disturbed. The experimental results conducted using the Pascal VOC2012 and Cityscapes data sets show that the suggested approach provides superior segmentation results than the current semi-supervised methods.