Underwater image compression for human and machine visions with hybrid priors embedding

The limited bandwidth in underwater acoustic communication requires the captured underwater images to be compressed and transmitted at extremely low bit rates. During the past decade, learning-based underwater image compression (UIC) methods have shown great potential. However, existing learning-based UIC methods mainly focus on human vision and do not fully consider the unique characteristics of underwater images. To tackle these challenges, this paper proposes a novel UIC method with Hybrid Priors Embedding (HPE) to guarantee the visual and analysis performance under low bit-rate and underwater degradation conditions. The proposed HPE-UIC method takes advantage of internal underwater physical priors to guide more compact bitrate allocation on the encoder side and external underwater statistical priors to facilitate the recovery of lost details due to underwater imaging distortions and compression-induced artifacts on the decoder side. Firstly, on the encoder side, we make use of the underwater imaging transmission map via an adaptive feature transformation module to enable a more compact representation of latent features. Secondly, on the decoder side, we construct a human-machine co-friendly underwater image feature codebook as the external high-quality prior to provide supplementary information for image reconstruction via a task-agnostic parallel decoder architecture. Due to the joint consideration of internal and external priors, the compression ratio is effectively improved and the quality of reconstructed image is well enhanced to meet the requirements of both human and machine vision tasks. Experimental results demonstrate the superiority of our HPE-UIC method in both human visual perception and machine analysis performance compared to existing UIC methods. © 2025 Elsevier Ltd.