《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (10): 3311-3319.DOI: 10.11772/j.issn.1001-9081.2024091383

• 多媒体计算与计算机仿真 • 上一篇    

基于盲环网络和随机恢复掩码的自监督图像去噪

梁震远1, 江松林2(), 朱松豪1   

  1. 1.南京邮电大学 自动化学院、人工智能学院,南京 210023
    2.闽西职业技术学院 城乡建筑学院,福建 龙岩 364021
  • 收稿日期:2024-09-27 修回日期:2025-01-08 接受日期:2025-01-10 发布日期:2025-03-18 出版日期:2025-10-10
  • 通讯作者: 江松林
  • 作者简介:梁震远(2000—),男,江苏南京人,硕士研究生,主要研究方向:图像处理、深度学习
    江松林(1972—),男,福建龙岩人,副教授,硕士,主要研究方向:电气自动化、图像处理 Email:331319315@qq.coom
    朱松豪(1973—),男,江苏如皋人,副教授,博士,主要研究方向:人工智能、深度学习。
  • 基金资助:
    国家自然科学基金资助项目(62001247)

Self-supervised image denoising based on blind-ring network and random recovery mask

Zhenyuan LIANG1, Songlin JIANG2(), Songhao ZHU1   

  1. 1.College of Automation and College of Artificial Intelligence,Nanjing University of Posts and Telecommunications,Nanjing Jiangsu 210023,China
    2.Urban and Rural Construction School,Minxi Vocational and Technical College,Longyan Fujian 364021,China
  • Received:2024-09-27 Revised:2025-01-08 Accepted:2025-01-10 Online:2025-03-18 Published:2025-10-10
  • Contact: Songlin JIANG
  • About author:LIANG Zhenyuan, born in 2000, M. S. candidate. His research interests include image processing, deep learning.
    JIANG Songlin, born in 1972, M. S., associate professor. His research interests include electric automation, image processing.
    ZHU Songhao, born in 1973, Ph. D., associate professor. His research interests include artificial intelligence, deep learning.
  • Supported by:
    National Natural Science Foundation of China(62001247)

摘要:

现有的基于盲点网络的自监督图像去噪方法常因为网络结构的限制,导致图像信息的严重损失。为解决这一问题,首先,提出一种自监督图像去噪方法,通过将传统的盲点网络改进为盲环网络(BRN),进一步降低噪声的空间相关性;其次,针对传统掩码策略导致图像信息丢失的问题,提出一种随机恢复掩码(RRM)策略,在减少信息损失的同时,增强去噪结果的细节信息;最后,提出一种双约束损失函数,在防止模型过度拟合的同时,有效保留图像的重要信息。实验结果表明,相较于次优的基于BRN的自监督图像去噪方法,所提方法在SIDD验证数据集上的峰值信噪比(PSNR)提高了0.17 dB,结构相似性(SSIM)提高了0.007,图像块感知相似度(IPPS)降低了0.006,验证了所提方法具有优越的去噪性能。

关键词: 图像去噪, 自监督学习, 盲环网络, 随机恢复掩码, 双约束损失函数

Abstract:

The existing self-supervised image denoising methods based on blind-spot networks often suffer from severe loss of image information due to limitations in the network structure. To solve this problem, firstly, a self-supervised image denoising method was proposed, which improved the traditional blind-spot network into a Blind-Ring Network (BRN), so as to further reduce spatial correlation of the noise. Then, to address the issue of image information loss caused by the traditional mask strategies, a Random Recovery Mask (RRM) strategy was proposed, thereby reducing the information loss while enhancing detail information of the denoising results. Finally, a dual constraint loss function was proposed to prevent over-fitting of the model while preserving important information in the image effectively. Experimental results show that compared with the sub-optimal self-supervised image denoising method based on BRN, the proposed method improves the Peak Signal-to-Noise Ratio (PSNR) by 0.17 dB, Structural SIMilarity (SSIM) by 0.007, and reduces the Image Patch Perceptual Similarity (IPPS) by 0.006, on the SIDD validation dataset, verifying its superior denoising performance.

Key words: image denoising, self-supervised learning, Blind-Ring Network (BRN), Random Recovery Mask (RRM), dual constraint loss function

中图分类号: