文献类型：期刊文献

英文题名：Quantization compensation and decomposition GAN for high-fidelity underwater image compression

作者：Hu, Xufei[1];Zhang, Jian[1];Zhang, Heng[1];Li, Ming[1];Huang, Meng[1];Liu, Jie[1];Zhang, Hengmin[2]

机构：[1]Jiangsu Ocean Univ, Sch Comp Engn, Lianyungang 222005, Peoples R China;[2]East China Univ Sci & Technol, Minist Educ, Sch Informat Sci & Engn, Key Lab Smart Mfg Energy Chem Proc, Shanghai 200237, Peoples R China

年份：2026

卷号：685

外文期刊名：NEUROCOMPUTING

收录：;EI(收录号：20261620506134);WOS:【SCI-EXPANDED(收录号:WOS:001765380600001)】；

基金：This work was supported by the Graduate Research and Practice Innovation Program of Jiangsu Province KYCX25_3859.

语种：英文

外文关键词：Underwater image compression; Latent space decomposition; Generative adversarial network; Rate-distortion performance

摘要：The increasing demand for marine exploration requires efficient transmission of underwater visual data, yet the limited bandwidth of acoustic channels makes ultra-low-bit-rate compression essential and challenging, espe cially for preserving structural and textural information. To address these issues, we propose a Quantization Compensation and Decomposition Generative Adversarial Network (QCD-GAN) for underwater image compres sion at ultra-low bit rates. First, to mitigate information loss in the quantized latent representation, a Quantization Compensation Generation Module (QCGM) and a Quantization Compensation Discrimination Module (QCDM) are introduced. Through adversarial learning, these modules reduce quantization-induced distortion and suppress blocking artifacts, thereby improving the naturalness of reconstructed images. Second, we introduce a latent space decomposition strategy, including a Structure Texture Separator (STS) and a Decomposer and Composer Fusion (DeCoF) module, to explicitly disentangle structure from fine-grained texture and thereby improve semantic fi delity and perceptual quality at ultra-low bit rates. Extensive experiments on four benchmark datasets show that the proposed QCD-GAN achieves superior performance. The proposed framework provides an effective solution for underwater visual data transmission in bandwidth and resource constrained environments, with potential applications in autonomous underwater vehicles and real-time deep-sea monitoring.