Multi-scale dense spatially-adaptive residual distillation network for lightweight underwater image super-resolution

Abstract

Underwater images are typically of poor quality, lacking texture and edge information, and are blurry and full of artifacts, which restricts the performance of subsequent tasks such as underwater object detection, and path planning for underwater unmanned submersible vehicles (UUVs). Additionally, the limitation of underwater equipment, most existing image enhancement and super-resolution methods cannot be implemented directly. Hence, developing a weightless technique for improving the resolution of submerged images while balancing performance and parameters is vital. In this paper, a multi-scale dense spatially-adaptive residual distillation network (MDSRDN) is proposed aiming at obtaining high-quality (HR) underwater images with odd parameters and fast running time. In particular, a multi-scale dense spatially-adaptive residual distillation module (MDSRD) is developed to facilitate the multi-scale global-to-local feature extraction like a multi-head transformer and enriching spatial attention maps. By introducing a spatial feature transformer layer (SFT layer) and residual spatial-adaptive feature attention (RSFA), an enhancing attention map for spatially-adaptive feature modulation is generated. Furthermore, to maintain the network lightweight enough, blue separable convolution (BS-Conv) and distillation module are applied. Extensive experimental results illustrate the superiority of MDSDRN in underwater image super-resolution reconstruction, which can achieve a great balance between parameters (only 0.32M), multi-adds (only 13G), and performance (26.38 dB on PSNR in USR-248) with the scale of ×4.