基于多特征融合和卷积神经网络的无参考图像质量评价  

No reference image quality assessment based on fusion of multiple features and convolutional neural network

作　　者：卢鹏[1] 刘楷贇 邹国良[1] 王振华[1] 郑宗生[1] LU Peng;LIU Kai-yun;ZOU Guo-liang;WANG Zhen-hua;ZHENG Zong-sheng(College of Information, Shanghai Ocean University, Shanghai 201306, China)

机构地区：[1]上海海洋大学信息学院,上海201306

出　　处：《液晶与显示》2022年第1期66-76,共11页Chinese Journal of Liquid Crystals and Displays

基　　金：上海市科技创新计划(No.20dz203800);上海市地方能力建设项目(No.19050502100);上海海洋大学科技发展专项(No.A2-2006-20-200211)。

摘　　要：为了更好地评价图像质量,解决在基于卷积神经网络的图像质量评价模型(CNN-IQA)上明显忽略的分块图像之间存在差异性的问题,提出了一种多特征融合的CNN模型。首先,将整幅图像进行不重叠分块,并提取每个分块图像的信息熵和纹理特征。然后,将提取计算的两特征相结合,计算各分块图像的重要性权重,以此衡量分块图像对失真图像质量的影响。最后,根据计算出的重要性权重修改损失函数,突出重要性高的分块图像在训练过程中发挥的作用。在LIVE数据集进行验证和对比实验发现,该算法的SROCC与LCC指标为0.962和0.960,相比原算法至少提升0.9%;在TID2008数据集上验证和对比实验发现,该算法获得的SROCC与LCC指标为0.922和0.926,相比原算法至少提升0.6%。并且在两数据集上的结果均优于其他对比算法。实验结果证明了其在预测图像质量方面具有良好的性能和泛化性。In order to evaluate the image quality better and solve the problem of the differences between the block images that are obviously ignored in the image quality evaluation model based on the convolutional neural network(CNN-IQA),a CNN model of multi-feature fusion is proposed.Firstly,the whole image is divided into non-overlapping blocks,the information entropy and texture feature of each divided image are extracted.Then,the two features are combined to calculate the importance weight of each block image to measure the influence of the block image on the quality of the distorted image.Finally,the loss function is modified according to the calculated importance weight to highlight the role of the block image with high importance in the training process.Validation and comparison experiments on the LIVE data set found that the SROCC and LCC indicators of the algorithm are 0.962 and 0.960,which are higher than the original algorithm at least 0.9%.The validation and comparison experiments on the TID2008 data set show that the SROCC and LCC indicators obtained by the algorithm are 0.922 and 0.924,which are higher than the original algorithm at least 0.6%.And the results on the two data sets are better than other comparison algorithms.The experimental results prove that it has good performance and generalization in predicting image quality.

关 键 词：图像处理 无参考图像质量评价 卷积神经网络 纹理特征 信息熵 损失函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]