Physical Prior-based Underwater Image Compression

Visual analysis of underwater images plays a pivotal role in underwater information exploration. Yet, it frequently encounters hurdles like image degradation and narrow communication bandwidth underwater, complicating detection and recognition in terminal systems. Current underwater image compression techniques, though fulfilling low bit-rate transmission needs, seldom factor in the distinctive physical characteristics of the underwater environment. This paper presents an innovative compression method for underwater images. By leveraging physical priors, it preprocesses degraded underwater images to enhance clarity, thus boosting the accuracy of machine vision analysis. The encoder employs a Transformer-based compression structure. In light of the physical traits and depth perception of underwater images, a Depth-Aware Feature Enhancement (DAFE) module is devised to enhance crucial information. Additionally, a prompt generation network is incorporated to allow for variable image quality, adapting to the restricted underwater acoustic channel bandwidth and striking a balance between reconstruction quality and compression ratio. Experiments confirm that the proposed framework shines in both machine vision tasks and perceptual quality.