Semi-supervised underwater image restoration based on dual-mean teachers and hybrid attention

Underwater image restoration is a fundamental technique for marine vision applications. Although supervised learning–based methods have achieved notable progress, their performance heavily relies on large-scale, high-quality paired datasets, which are difficult to acquire in real-world underwater environments. Semi-supervised learning, particularly Mean Teacher–based frameworks, partially alleviates this limitation by leveraging unlabeled data. However, existing methods still suffer from limited supervision diversity caused by single-teacher pseudo labeling and insufficient capability to jointly model global color distortion and local detail degradation. To address these issues, we propose a Dual Collaborative Mean Teacher (DCMT) framework for semi-supervised underwater image restoration. DCMT introduces a homogeneous dual-teacher architecture, in which two teacher models are independently maintained through the exponential moving average (EMA) updates from two separately optimized student networks. This design enhances supervision diversity by coupling each teacher with a different student trajectory, thereby reducing the risk of pseudo-label error accumulation and providing more robust consistency guidance. In addition, a reliability-aware pseudo-label selection mechanism is introduced to conservatively filter teacher predictions according to relative quality assessment. We also employ a SwinCBAM-based hybrid attention design as an effective combination of global context modeling and local feature refinement, enabling joint modeling of global color correction and local detail restoration. Extensive experiments on multiple benchmark datasets show that the proposed method achieves strong quantitative and visual performance, with consistent improvements in color fidelity and detail preservation.