Progressive Attentional Learning for Underwater Image Super-Resolution

Visual perception plays an important role when underwater robots carry out missions under the sea. However, the quality of images captured by visual sensors is often affected by underwater environment conditions. Image super-resolution is an effective way to enhance the resolution of underwater images. In this paper, we propose a novel method for underwater image super-resolution. The proposed method uses CNNs with channel-wise attention to learn a mapping from low-resolution images to high-resolution images. And a progressive training strategy is used to deal with large scaling factors (e.g. 4x and 8x) of super-resolution. We name our method as Progressive Attentional Learning (PAL). Experiments on a recently published underwater image super-resolution dataset, USR-248 [11], show the superiority of our method over other state-of-the-art methods.