Underwater Image Enhancement Based on Transformer, Attention, and Multi-Color-Space Inputs

Underwater images suffer from color distortion, low contrast and blurring caused by the attenuation, refraction, and scattering of light. For many maritime operations, underwater image enhancement is essential. This paper proposes an underwater image enhancement method based on transformer, attention and multi-color-space inputs. First, transformer block is embedded into ResNet-50 as the backbone network. Combining this backbone network and multi-scale feature fusion forms the fundamental framework of our method. This framework contributes to mine the rich features of different scene underwater image and outputs a feature map of the same size as the input image. Then, multi-dimensional attention mechanism is added to the feature maps of different scales to mine key areas that require enhancement in underwater images and calculate the enhancement value of each pixel. Finally, multi-color-space inputs including RGB, HSV and LAB as the inputs for CNN to enhancement images from various aspects such as color deviation, brightness and saturation. The comparison of image quality evaluation metrics, visual enhancement performance evaluation and impact on the performance of underwater object detection with other underwater enhancement methods on datasets including SUIM-E, UIEB, EUVP, RUIE-RIQS, RUIE-UCCS and URPC-2018 prove the good performance of our underwater image enhancement method.