Underwater-image super-resolution via range-dependency learning of multiscale features

Optical imaging instruments have been widely deployed in underwater-engineering systems, playing an important role in underwater-object localization, detection, and recognition. However, underwater images suffer from the adverse effects of local distortion and global haze. Consequently, raw underwater images have limitations when used for display and vision tasks. Super resolution (SR) has been increasingly exploited for perceptual image quality improvement. However, underwater-image SR is a relatively underexplored area, most existing methods being unable to reduce the adverse effects of underwater images. Moreover, in contrast to land-based high-performance platforms, the power supply and computational resources of underwater platforms are limited, making them difficult to use in large-scale models. To solve these problems, this study developed a novel range-dependency learning network to present the short- and long-range dependency of multiscale features. Such a mechanism could provide more detailed and accurate texture information for underwater-image SR, improving underwater-image SR performance. Moreover, a channel-splitting module was designed to generate the channel bands which could extract texture details and global structural information at different scales while reducing the number of parameters, thus accelerating the training speed of the model and maintaining good performance. Our novel network could reach an optimal tradeoff between the underwater-image SR performance and efficiency, which was demonstrated by experimental comparisons and an ablation study.