Abstract
Semi-supervised learning for change detection (CD), which significantly reduces the labor costs associated with data annotation, has recently garnered substantial attention. In this study, we propose to enhance traditional semi-supervised learning frameworks by leveraging cross-task cross-domain (CTCD) models, which generate complementary features that differ from standard hidden features. The procedure is as follows. Firstly, the standard features obtained from a traditional encoding-decoding structure are fused with attention-augmented complementary features. Secondly, a secondary decoder maps the fused heterogeneous features into the label space to obtain high-quality pseudo-labels, offering more precise guidance for semi-supervised learning on traditional structures. This approach improves pseudo-labels by leveraging the strength of CTCD models, including large pre-trained models, to enhance the semi-supervised learning process of domain-specific and task-specific models. Experimental results on benchmark datasets demonstrate that our proposed approach surpasses state-of-the-art methods.
Funding Information
  • Ningbo Science and Technology Innovation Project (2023CX050006)
  • Ningbo Science and Technology Innovation Project (2022Z075)