ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2019, Vol. 56 ›› Issue (11): 2458-2468.doi: 10.7544/issn1000-1239.2019.20180617

• 图形图像 • 上一篇    下一篇

图像质量感知的混合噪声快速盲降噪算法

徐少平,刘婷云,罗洁,张贵珍,唐祎玲   

  1. (南昌大学信息工程学院 南昌 330031) (xushaoping@ncu.edu.cn)
  • 出版日期: 2019-11-12
  • 基金资助: 
    国家自然科学基金项目(61662044,61163023,51765042);江西省自然科学基金项目(20171BAB202017)

An Image Quality-Aware Fast Blind Denoising Algorithm for Mixed Noise

Xu Shaoping, Liu Tingyun, Luo Jie, Zhang Guizhen, Tang Yiling   

  1. (School of Information Engineering, Nanchang University, Nanchang 330031)
  • Online: 2019-11-12

摘要: 现有的高斯-脉冲混合噪声降噪算法多基于正则化技术采用迭代求解最优目标函数值的方式实现,执行效率普遍比较低,严重限制了其实际应用范围.为此,以卷积神经网络(convolutional neural network, CNN)为核心技术提出了一种基于图像质量感知的快速盲降噪算法(image quality-aware fast blind denoising algorithm, IQA-FBDA).在训练阶段,首先基于浅层CNN卷积神经网络设计图像质量评估模型来预测待降噪图像的图像质量值;然后,依据在大量噪声图像训练集合上获得的图像质量值统计分布规律构建混合噪声模式分类字典;最后,基于该分类字典将噪声图像集合划分为16个子集并训练与各个子集相匹配的深层CNN卷积神经网络专用降噪模型.在降噪阶段,首先利用图像质量评估模型估计给定待降噪图像的质量值,然后依据所预测的图像质量值查找噪声模式分类字典并调用与之相匹配预先训练好的深层CNN降噪模型即可快速地完成盲降噪任务.实验数据表明:IQA-FBDA算法在降噪效果方面的性能达到了与主流高斯-脉冲混合噪声降噪算法相当的水平,而在执行效率方面则有极大提高,更具实用价值.

关键词: 图像降噪, 高斯-脉冲混合噪声, 图像质量感知, 噪声模式分类字典, 卷积神经网络, 执行效率

Abstract: The existing Gaussian-impulse mixed noise removal algorithms usually restore the noisy images via regularization technique by solving an optimal objective function iteratively, which results in low executive efficiency and limits their practical applications. To this end, in this paper we propose an image quality-aware fast blind denoising algorithm (IQA-FBDA), which takes convolutional neural network (CNN) as the core technique for the removal of Gaussian-impulse mixed noise. In the training phase, a shallow CNN-based image quality estimation model is first exploited to estimate the image quality of the image to be denoised. Then, according to the statistical distribution of the image qualities of a large number of noisy images, we construct a mixed noise pattern classification dictionary (MNPCD). Based on the MNPCD, the training noisy images are classified into 16 sub-classes, and then deep CNN-based denoisers for each class are trained. In the denoising phase, the image quality estimation model is first used to estimate the quality value of a given noisy image. After querying the quality value in the MNPCD, the corresponding pre-trained denoiser is exploited to achieve efficient blind image denoising. Experiments show that, compared with the state-of-the-art Gaussian-impulse mixed noise removal algorithms, the proposed one achieves comparable noise reduction effect with great improvement in terms of the execution efficiency, which makes it more practical.

Key words: image denoising, Gaussian-impulse mixed noise, image quality-aware, noise pattern classification dictionary, convolutional neural network (CNN), execution efficiency

中图分类号: