基于对比学习的图像压缩感知预训练方法  

Pretraining Method for Image Compressive Sensing Based on Contrastive Learning

作　　者：许常宜 丁海峰 王海栋 XU Changyi;DING Haifeng;WANG Haidong(Suzhou Rail Transit Group Co.,Ltd.,Suzhou Jiangsu 215008,China;Suzhou Rail Transit Group Operating Co.,Ltd.,Suzhou Jiangsu 215008,China;Department of Mathematics,Soochow University,Suzhou Jiangsu 215006,China)

机构地区：[1]苏州轨道交通集团有限公司,江苏苏州215008 [2]苏州轨道交通集团运营有限公司,江苏苏州215008 [3]苏州大学数学科学学院,江苏苏州215006

出　　处：《微处理机》2025年第2期14-20,共7页Microprocessors

基　　金：江苏省重点研发项目(No.BE2022058-4)。

摘　　要：针对当前基于深度学习的图像压缩感知方法中非线性采样算子预训练方案不足的问题,提出一种基于对比学习的预训练方法。该方法从限制等距性质(RIP)矩阵的特性出发,利用对比学习优化采样算子,使其能够更有效地捕捉样本间的相似性关系,从而在采样阶段保留更多图像信息。通过将对比学习引入非线性采样算子的预训练过程,增强了采样算子的表征能力。实验表明该方案显著提升了图像重构质量,同时加快了模型收敛速度,提高了训练效率。基于对比学习的预训练策略能够有效优化非线性采样算子,为图像压缩感知的性能提升提供了新的思路。To address the lack of effective pretraining schemes for nonlinear sampling operators in deep learning-based image compressive sensing,this paper proposes a contrastive learning-based pretraining method.Leveraging the properties of the Restricted Isometry Property(RIP)matrix,the proposed approach optimizes the sampling operator through contrastive learning,enabling it to better capture inter-sample similarity relationships and thus preserve more image information during sampling.By integrating contrastive learning into the pretraining process of nonlinear sampling operators,the representational capability of the sampling operator is enhanced.Experimental results demonstrate that the proposed method significantly improves image reconstruction quality while accelerating model convergence and enhancing training efficiency.The contrastive learning-based pretraining strategy effectively optimizes nonlinear sampling operators,offering a new approach for improving the performance of image compressive sensing.

关 键 词：图像压缩感知 对比学习 预训练 非线性采样算子 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]