Domain adaptive underwater object detection via complementary style-aware learning.

Abstract

Images from different underwater domains exhibit multi-style variations caused by water quality, lighting, and imaging devices. The domain gap caused by these variations, combined with limited annotated data in target domains, leads to degraded object detection performance in cross domain situations. Mean Teacher is an effective framework to address performance degradation caused by cross-domain discrepancies, but its effectiveness is constrained by the quality of pseudo-labels generated by the teacher model. To address this, we propose a complementary style-aware Mean Teacher (CSAMT) model. It constructs image pairs to facilitate style-aware learning, performs style-content disentanglement via WCT2, and leverages Discrete Wavelet Transform (DWT)'s band-separation properties to jointly model image-pair features across frequency and spatial domains. Additionally, a two-stage teacher-student region proposal alignment (TTRPA) strategy is introduced, which guides the model to assign higher attention weights to more effective regions and constructs a consistency loss for suboptimal supervision. Experiments across three domain adaptation benchmarks demonstrate state-of-the-art performance, with ablation studies validating the effectiveness of each component.