结构-语义-残差协同的图像压缩框架：面向隐私保护与高保真重建

A Structure-Semantics-Residual Collaborative Image Compression Framework: Towards Privacy Protection and High-Fidelity Reconstruction

摘要

为在极低码率条件下同时实现图像隐私保护与高保真重建，提出了一种结构-语义-残差协同的图像压缩框架。该方法从图像结构、语义与残差3个层面进行表征提取：利用边缘图作为结构信息进行主干压缩；提取去互信息处理的图像文本描述作为语义先验进行指导；将包含敏感区域信息的残差信号通过可信通道传输。在解码端，一方面结合边缘图与语义信息生成去隐私化图像，另一方面利用边缘图与残差信息恢复高保真原始图像，从而实现隐私安全与视觉保真的协同优化。结果表明：在CLIC数据集上的对比实验中，所提方法在实现高保真图像重建的同时，可节省13%～32%的码率，并能够分路生成去隐私化图像；消融实验验证了各个分支结构及约束机制的有效性，表明该框架能够在隐私保护与图像重建质量之间实现合理权衡。

Abstract

To achieve both image privacy protection and high-fidelity reconstruction at ultra-low bitrates, a structure-semantics-residual collaborative image compression framework is proposed. In this framework, representations are extracted from three distinct levels (i.e., image structure, semantics and residual): structural information is captured via edge maps for backbone compression; semantic priors are derived from image captions with mutual information removed; and residual signals containing sensitive region information are transmitted via a trusted channel. At the decoder, de-identified images are generated by integrating edge maps with semantic information, while original high-fidelity images are restored using edge maps and residual signals, thereby achieving synergistic optimization of privacy protection and visual fidelity. Results demonstrate that through the comparative experiments on the CLIC (Challenge on Learned Image Compression) dataset, the proposed method reduces bitrate by 13%—32% while maintaining high-quality reconstruction. Additionally, branch-wise generation of de-identified images is supported. Furthermore, the effectiveness of the individual branches and constraint mechanisms is verified through ablation experiments, indicating that a reasonable trade-off between privacy protection and image reconstruction quality is realized by the proposed framework.

关键词

图像压缩 / 极低码率 / 隐私保护 / 高保真

Key words

image compression / ultra-low bitrate / privacy protection / high-fidelity

参考文献

[1] Wallace G K . The JPEG still picture compression standard[J].Communications of the ACM, 1991, 34(4): 30-44.

[2] Christopoulos C, Skodras A, Ebrahimi T . The JPEG2000 still image coding system: an overview[J].IEEE Transactions on Consumer Electronics, 2000, 46(4): 1103-1127.

[3] Sullivan G J, Ohm J R, Han W J, et al. Overview of the high efficiency video coding (HEVC) standard[J].IEEE Transactions on Circuits and Systems for Video Technology, 2012, 22(12): 1649-1668.

[4] 宋蓓蓓，何帆，马穗娜，等 . 一种高精度的自适应码率控制图像压缩算法[J].西安交通大学学报，2022，56(2): 198-206.

[5] Song Beibei, He Fan, MA Suina, et al. An image compression algorithm with high—precision adaptive rate control[J].Journal of Xi’an Jiaotong University, 2022, 56(2): 198-206.

[6] Ballé J, Laparra V, Simoncelli E P . End—to—end optimized image compression[EB/OL]. (2017—03—03) [2025—07—27].https://arxiv.org/abs/1611.01704.

[7] Cheng Zhengxue, Sun Heming, Takeuchi M, et al. Learned image compression with discretized Gaussian mixture likelihoods and attention modules[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, NJ, USA: IEEE, 2020: 7936-7945.

[8] Iwai S, Miyazaki T, Sugaya Y, et al. Fidelity—controllable extreme image compression with generative adversarial networks[C]//2020 25th International Conference on Pattern Recognition (ICPR). Piscataway, NJ, USA: IEEE, 2021: 8235-8242.

[9] Agustsson E, Tschannen M, Mentzer F, et al. Generative adversarial networks for extreme learned image compression[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway, NJ, USA: IEEE, 2019: 221-231.

[10] Lei E, Uslu Y B, Hassani H, et al. Text+sketch: image compression at ultra low rates[EB/OL]. (2023—06—01) [2025—07—27].https://arxiv.org/abs/2306.00040.

[11] Li C, Lu G, Feng D, et al. MISC: ultra—low bitrate image semantic compression driven by large multimodal model[J].IEEE Transactions on Image Processing, 2025, 34(1): 335-349.

[12] 张云飞，贺丽君，王子溪，等 . 面向机器视觉任务的多尺度图像特征压缩算法[J].西安交通大学学报，2023，57(12): 1-10.

[13] Zhang Yunfei, He Lijun, Wang Zixi, et al. Image multi—scale feature compression algorithm for machine vision tasks[J].Journal of Xi’an Jiaotong University, 2023, 57(12): 1-10.

[14] 马彩霞，贾春福，蔡智鹏，等 . 面向对抗样本干扰的人脸识别隐私保护方案[J].西安电子科技大学学报，2025，52(2): 190-200.

[15] Ma Caixia, Jia Chunfu, Cai Zhipeng, et al. Privacy—preserving face recognition against adversarial sample perturbations[J].Journal of Xidian University, 2025, 52(2): 190-200.

[16] 周浩，戴华，杨庚，等 . 基于生物特征识别的隐私保护可验证联邦学习[J].计算机学报，2025，48(8): 1848-1869.

[17] Zhou Hao, Dai Hua, Yang Geng, et al. Privacy—preserving verifiable federated learning based on biometric recognition[J].Chinese Journal of Computers, 2025, 48(8): 1848-1869.

[18] Rakhmawati L . Image privacy protection techniques: a survey[C]//TENCON 2018: 2018 IEEE Region 10 Conference. Piscataway, NJ, USA: IEEE, 2018: 0076-0080.

[19] Yu Jun, Zhang Baopeng, Kuang Zhengzhong, et al. iPrivacy: image privacy protection by identifying sensitive objects via deep multi—task learning[J].IEEE Transactions on Information Forensics and Security, 2017, 12(5): 1005-1016.

[20] 闫光辉，刘婷，张学军，等 . 抵御背景知识推理攻击的服务相似性位置k匿名隐私保护方法[J].西安交通大学学报，2020，54(1): 8-18.

[21] Yan Guanghui, Liu Ting, Zhang Xuejun, et al. Service similarity location k anonymity privacy protection scheme against background knowledge inference attacks[J].Journal of Xi’an Jiaotong University, 2020, 54(1): 8-18.

[22] Liu Chi, Zhu Tianqing, Zhang Jun, et al. Privacy intelligence: a survey on image privacy in online social networks[J].ACM Computing Surveys, 2023, 55(8): 161.

[23] 李璇，李德华，杨智，等 . 图像中人脸隐私度的定量评估研究[J].计算机与数字工程，2019，47(10): 2550-2555.

[24] Li Xuan, Li Dehua, Yang Zhi, et al. Quantifying privacy levels of faces in images[J].Computer & Digital Engineering, 2019, 47(10): 2550-2555.

[25] Shan S, Wenger E, Zhang Jiayun, et al. Fawkes: protecting privacy against unauthorized deep learning models[C]//Proceedings of the 29th USENIX Conference on Security Symposium. USA: USENIX Association, 2020: 1589-1604.

[26] 周治平，钱新宇 . 一种面向深度神经网络的差分隐私保护算法[J].电子与信息学报，2022，44(5): 1773-1781.

[27] Zhou Zhiping, Qian Xinyu . Differential privacy algorithm under deep neural networks[J].Journal of Electronics & Information Technology, 2022, 44(5): 1773-1781.

[28] Zhang Yaofang, Fang Yuchun, Cao Yiting, et al. RB—GAN: realistic—generation and balanced—utility GAN for face de—identification[J].Image and Vision Computing, 2024, 141: 104868.

[29] Wu Yifan, Yang Fan, Ling Haibin . Privacy—protective—GAN for face de—identification[EB/OL]. (2018—06—23) [2025—07—27].https://arxiv.org/abs/1806.08906.

[30] Ruchaud N, Dugelay J L . JPEG—based scalable privacy protection and image data utility preservation[J].IET Signal Processing, 2018, 12(7): 881-887.

[31] Zhang Bo, Xiao Di, Li Xiuling, et al. Privacy—preserving image compressed sensing by embedding a controllable noise—injected transformation for IoT devices[J].Signal Processing, 2023, 210: 109055.

[32] Xie Saining, Tu Zhuowen . Holistically—nested edge detection[C]//2015 IEEE International Conference on Computer Vision (ICCV). Piscataway, NJ, USA: IEEE, 2015: 1395-1403.

[33] Bianchi F, Attanasio G, PISONI R, et al. Contrastive language—image pre—training for the Italian language[EB/OL]. (2021—08—19) [2025—07—27].https://arxiv.org/abs/2108.08688.

[34] Poole B, Ozair S, van den Oord A, et al. On variational bounds of mutual information[C]//Proceedings of the 36th International Conference on Machine Learning. Chia Laguna Resort, Sardinia, Italy: PMLR, 2019: 5171-5180.

[35] Belghazi M I, Baratin A, Rajeswar S, et al. MINE: mutual information neural estimation[EB/OL]. (2021—08—14) [2025—07—27].https://arxiv.org/abs/1801.04062.

[36] Ballé J, Chou P A, Minnen D, et al. Nonlinear transform coding[J].IEEE Journal of Selected Topics in Signal Processing, 2021, 15(2): 339-353.

[37] Ronneberger O, Fischer P, Brox T . U—Net: convolutional networks for biomedical image segmentation[C]//Medical Image Computing and Computer—Assisted Intervention—MICCAI 2015. Cham, Germany: Springer International Publishing, 2015: 234-241.

[38] Simonyan K, Zisserman A . Very deep convolutional networks for large—scale image recognition[EB/OL]. (2015—04—10) [2025—07—27].https://arxiv.org/abs/1409.1556.

[39] Bellard F. BPG image format[EB/OL]. [2025—07—26].https://bellard.org/bpg/.

[40] ISO/IEC 23090—3:2024 Information technology—coded representation of immersive media: part 3 versatile video coding[S].

[41] Fu Haisheng, Liang Feng, Lin Jianping, et al. Learned image compression with Gaussian—Laplacian—logistic mixture model and concatenated residual modules[J].IEEE Transactions on Image Processing, 2023, 32: 2063-2076.

基金资助

国家自然科学基金资助项目(62471376)