Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (21): 243-249.DOI: 10.3778/j.issn.1002-8331.2103-0560

• Graphics and Image Processing • Previous Articles     Next Articles

DCN:Image Quality Assessment Network of Dual-Channel Dense Hadamard Product

YANG Xiaodong, HAN Zhenqi, LIU Lizhuang, ZHAO Dan   

  1. 1.Shanghai Advanced Research Institute, Chinese Academy of Sciences, Shanghai 201210, China 
    2.University of Chinese Academy of Sciences, Beijing 100049, China
  • Online:2022-11-01 Published:2022-11-01



  1. 1.中国科学院 上海高等研究院,上海 201210
    2.中国科学院大学,北京 100049

Abstract: Image quality assessment(IQA) is becoming more and more important as the human needs for high quality images become more and more urgent. The non-reference authentic distortion assessment is faced with great challenges due to the complexity of distortion and content diversity. In order to obtain more accurate and effective quality characteristics, the dual-channel dense Hadamard convolution image quality evaluation network(DCN) is proposed. In DCN, the deep convolution model of Infection-ResNet-V2 is used as the backbone network to extract features, and the designed dual-channel converged network is used as the score evaluation network, which is finally mapped to the objective quality score. The score evaluation network is composed of a convolution feature extraction branch and a multilayer perceptron branch in parallel. In the multilayer perceptron branch, a dense hadamard product modul(DHPM) is proposed. Through the Hadamard product, low-level features and high-level features are integrated to play the role of feature adaptation and high-level expression. Experimental results on the open data set Koniq-10K show that the Spearman rank order correlation coefficient(SROCC) and Pearson linear correlation coefficient(PLCC) of the network test are 0.922 and 0.938 respectively.

Key words: image quality assessment, deep learning, computer vision, dual-channel structure, Hadamard product

摘要: 随着人类对高质量图像的需求日益紧迫,客观画质评价(image quality assessment,IQA)的研究日趋重要,其中的无参考真实失真评估,面临失真的复杂性和内容多样性的巨大挑战。为了获取更加准确有效的质量特征,提出了一种双通道密集哈达玛卷积的画质评价网络(dual-channel network,DCN),其以深度卷积模型Inception-ResNet-v2为骨干网络提取特征,将设计的双通道融合网络为分数评估网络,最后映射到客观质量分数。分数评估网络由卷积特征提取分支和多层感知机分支并联组成,将提出的密集哈达玛卷积模块(dense Hadamard product module,DHPM)应用到多层感知机分支中,通过哈达玛乘积将低层特征与高层特征融合,发挥特征自适应和高级表达的作用。在公开数据集KonIQ-10k上的实验结果表明,该网络测试的斯皮尔曼秩相关系数(spearman rankorder correlation coefficient,SROCC)为0.922,皮尔森线性相关系数(Pearson linear correlation coefficient,PLCC)达到0.938。

关键词: 图像质量评价, 深度学习, 计算机视觉, 双通道结构, 哈达玛卷积