计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (9): 26-32.DOI: 10.3778/j.issn.1002-8331.1812-0351

• 热点与综述 • 上一篇    下一篇

双目标的CNN无参考图像质量评价方法

程晓梅,沈远彤   

  1. 中国地质大学(武汉) 数学与物理学院,武汉 430074
  • 出版日期:2019-05-01 发布日期:2019-04-28

Double-Target CNN Image Quality Assessment Method

CHENG Xiaomei, SHEN Yuantong   

  1. School of Mathematics and Physics, China University of Geosciences, Wuhan 430074, China
  • Online:2019-05-01 Published:2019-04-28

摘要: 为有效提取与人类视觉感知高度相关的图像质量特征,提出了一个估计图像退化类型和质量评分的双目标卷积神经网络(CNN)结构。该网络结构有次序地分步提取用作退化类型分类和用作估计质量评分的特征,使网络更充分地挖掘图像退化类型信息并强化其对质量评分估计任务的辅助作用,进而提升了网络对图像质量特征的学习能力,同时实验表明两步特征提取的方式能加速网络的收敛。通过在标准图像质量评价数据库LIVE和TID2008上的对比实验,结果表明该算法在图像退化类型和质量评分两个任务中,整体性能均明显优于其他经典评价方法。

关键词: 无参考图像质量评价, 双目标, 卷积神经网络, 特征学习, 次序

Abstract: In order to effectively extract the image quality features highly correlated with human visual perception, the double-target convolutional neural network is proposed in this paper, which can estimate image distortion type and quality scores. The network structure trained the distortion type features and quality features of the image sequentially to make the network more fully excavate the image degradation type information and strengthen its auxiliary role in the quality score estimation task, and then improve its learning ability for image quality features. Simultaneously, the experiment indicated the two-step feature extraction method can accelerate the convergence of the network. Comparison experiments are carried out on the standard image quality evaluation database LIVE and TID2008, which show the algorithm can accurately evaluate the image quality scores and recognition distortion types, obviously better than other evaluation methods.

Key words: no-reference image quality assessment, double-target, Convolutional Neural Network(CNN), feature learning, sequentially