Building Deeper with U-Attention Net for Underwater Image Enhancement

The images captured in the underwater scene suffer from color casts due to the scattering and absorption of light. These problems severe interfere many vision tasks in the underwater scene. In this paper, we propose a Deeper U-Attention Net for underwater image enhancement. Different from most existing underwater image enhancement methods, we adequately exploit underlying complementary information of different scales, which can enrich the feature representation from multiply perspectives. Specifically, we design a novel module with self-attention module to enhance the features by the features itself. Then, we design U-Attention block to extract the features at a certain level. At last, we build deeper two level U-structure net with the proposed U-attention block at multiply scale. This architecture enables us to build a very deep network, which can extract the multi-scale features for underwater image enhancement. Experimental results show our method has a better performance on public datasets than most state-of-the-art methods.