融入注意力机制的弱监督水下图像增强算法

The supervised underwater image enhancement algorithms need paired training image samples that are difficult to be obtained in some uncontrolled scenarios such as underwater scenarios. A weakly supervised underwater image enhancement algorithm incorporating attention mechanism was proposed. Firstly, the red channel attenuation map was calculated according to the characteristics that the light with different wavelengths suffers from different attenuation when it propagates in water. After that, the attention module guided by the calculated red channel attenuation map was integrated into the generator, which effectively improved the performance of the generator in terms of correcting the color deviation of underwater images. In addition, a multiple joint loss function, including an adversarial loss and a structural similarity loss, was designed, which retained more image details while correcting color deviation of underwater images. Finally, the underwater image enhancement network was optimized under global and local scales. Experimental results show that the proposed algorithm is better than the competing algorithms in both subjective visual quality and objective evaluation index, and thus can effectively improve the visibility of underwater images.