生成对抗网络复原图像的质量测评方法研究

Research on Quality Evaluation Method for Restored Image Based on Generative Adversarial Network

针对提高基于生成式对抗网络的复原图像可信度的需求,提出一种基于多方向差异哈希(MDhash)算法的复原图像质量测评方法。该方法基于像素域,根据相邻像素之间灰度值的差异计算图像的哈希值。通过组合横向、纵向、斜向上、斜向下四个方向的Hash值得到多方向差异哈希值,实现图像的准确表征。通过计算四个方向上复原图像与参考图像的MDhash值之间的汉明距离,度量复原图像与原始图像的相似性。最后,基于公开数据集对所提评价指标与主观评分值的相关性进行验证。实验结果表明,本方法可实现对复原图像质量的准确量化评估,与对比方法结构相似度(SSIM)相比,MDhash方法与人的主观评价结果具有更高的相关系数,相关系数为0.9819。

To improve the reliability of the restored image based on generative adversarial network,a restored image quality evaluation method on multi⁃direction differential hashing(MDhash)algorithm is proposed for military infrared images.Based on the pixel domain,the method calculates the image hash value according to the difference of gray value between adjacent pixels.Firstly,by combining four hashes in horizontal,vertical,oblique upward and oblique downward directions,the hash value of multi⁃direction difference can be obtained,and the accurate representation of the image can be achieved.Then the hamming distance between the restored image and the reference image in four directions is calculated to measure the similarity between them.Finally,the correlation between the proposed evaluation indicators and the subjective score values can be verified with the open dataset.Experimental results show that this method can achieve accurate quantitative evaluation of the restored image quality.Compared with structural similarity(SSIM)of the contrast method,the proposed MDhash method has a higher correlation coefficient with the subjective results of human,and the correlation coefficient is 0.9819.