Abstract
Underwater image super-resolution reconstruction technologies have played a very important role in ocean resource exploration since it can significantly improve the clarity of underwater optical images. Although recent deep learning based methods have achieved promising performance in terrestrial image super-resolution, these methods lack sufficient capabilities to handle those dark, turbid and blurred underwater images. In this work, we propose a novel network, namely feature separable reconstruction network (FSRN), to separate the extraction features and the reconstruction features for better using of features in each layer, solving the problem of long-distance transmission of shallow features in the neural network. We design a depthwise separable convolutional residual block with large convolutional kernels (DWRB) to augment receptive fields, which improves the effectiveness of high-frequency feature extraction in the blur images. We further propose a channel attention mechanism based on the SE module and explore an optimal attention module insertion mode which pays more attention to the weight between reconstruction information, reducing information loss. Moreover, we also modify the convolutional kernel padding mode and propose a perceptual loss function with boundary clipping to avoid the inconsistent in feature extraction from boundary and non-boundary regions. Extensive experiments on underwater datasets demonstrate our proposed underwater super-resolution framework outperform over the state-of-the-art methods in terms of reconstruction accuracy and real-time performance.
