For the stable operation of marine robots, a robust visual perception system is required that can function reliably even under adverse visual conditions such as diffuse reflection from the sea surface, wave foam, underwater turbidity, and changes in i...
For the stable operation of marine robots, a robust visual perception system is required that can function reliably even under adverse visual conditions such as diffuse reflection from the sea surface, wave foam, underwater turbidity, and changes in illumination. This paper proposes a reinforcement method for a deep learning-based visual perception model to ensure stable recognition performance even under such adverse conditions. First, an Object-Centric Attention Module(OCAM) is proposed to suppress strong visual noise originating from the sea surface and effectively highlight the features of maritime objects. This reduces background interference such as sea surface reflections and wave foam, thereby improving detection performance for actual objects. Second, to enhance object segmentation accuracy in underwater environments with low contrast due to high turbidity, a weighted loss function is proposed that assigns greater learning weight to object boundary regions. This method is designed to enable stable learning of structural shapes even with blurred boundaries. The proposed enhancement methods were validated using real-world field data in UAV-based maritime search and rescue (SAR) and ROV-based aquaculture net damage detection scenarios. They demonstrated improved model recognition stability and reliability even under various adverse visual conditions.