基于多层次特征增强的图像压缩感知重构算法

打开文本图片集

关键词：图像压缩感知；动态偏移自注意力；多层次特征增强；双分支架构 中图分类号：TP391.4 文献标志码：A 文章编号：1001-3695（2026）04-038-1272-09 doi：10.19734/j.issn.1001-3695.2025.07.0289

Image compressive sensing reconstruction algorithm based on multilevel feature enhancement

Liu Yuhong，Li Dan† （SchoolofElectronicandInformation Engineering，Lanzhou Jiaotong University，Lanzhou 73oo7O，China）

Abstract：Thetraditional imagecompressivesensing reconstruction algorithmsoftensufferfrom limitations indetailedlocal featuremodeling.Transformers with globalself-atentioncancaptureglobal featuresbutsufferfromhighcomputationalcomplexity.Window-basedself-atentionmechanismreducescomplexity，butitslimitedcross-indowinteractionrestrictsthe modelingoflong-rangedependencies.Consequently，thispaperproposedamulti-level featureenhancementnetwork （MFENet）.In the sampling phase，itacquired more efective measurements bylearning asampling matrix.During the reconstructionphase，itdesignedadynamicoffsetself-attentionmechanismthatadaptivelyadjustedatentionheadsamplingpositionstoenhanceinter-windowinformationinteraction，therebyimprovingtheglobalfeaturemodelingcapabilityofTransformers.Byintegrating thelocal featureextractionadvantagesofCNNs，itconstructedadual-brancharchitecture tooptimize multi-levelfeatures.Furthermore，itachievedend-to-endmodeltraningthroughjointoptimizationofsamplingandreconstruction.Experimental results demonstrate that，on the Set11 test set at a 10% sampling rate，the peak signal-to-noise ratio and structuralsimilarityimprovebyO.53dBandO.Oo51，respectively，comparedtoCSformer.Theaveragereconstructiontime across all sampling rates decreases by 2.897 8 s，proving the algorithm’s effectiveness.

Key words：imagecompressvesensing（ICS）；dynamicofsetself-attention；multi-level featuresenhancement；dual-branch architecture

0 引言

压缩感知（compressivesensing，CS）理论表明，当信号可以进行稀疏表示时，能够以远低于奈奎斯特采样定律的采样率对其进行采样，得到CS测量值，并高概率地恢复出原信号[1]。（剩余21220字）