基于注意力和最小可觉差的GAN生成图像质量评价

GAN-generated image quality assessment based on attentionand just noticeable difference

现有的客观图像质量评价方法用于GAN生成图像质量评价时,往往出现与人的主观评价不一致的情况.针对这个问题,提出了一种更符合人类视觉感知的GAN生成图像质量客观评价方法AJ-GIQA(attention and just noticeable difference based generated image quality assessment).首先,模拟人类视觉系统的失真敏感度特性,对GAN生成图像进行预处理,得到其最小可觉差图;然后,将注意力模块引入特征提取网络,模拟人类视觉系统的注意力机制,获取图像的显著性特征;最后,将图像特征输入结合语义信息的质量预测网络,基于图像内容综合评价GAN生成图像的质量.在GAN生成图像数据集上的实验结果表明,AJ-GIQA的评价结果与主观平均意见得分有更高的一致性;在图像质量排序一致性上的实验结果表明,AJ-GIQA的准确率在LGIQA-LSUN-cat数据集上达到了最优,和SFA方法相比性能提高了0.267;在泛化性能上的实验结果表明,与最先进的HyperIQA方法相比,AJ-GIQA在数据集PIPAL的Pearson线性相关系数提高了0.027.

When the objective image quality assessment methods are used to assess the quality of GAN-generated images,the results are often inconsistent with human’s subjective evaluation.In order to solve this problem,this paper proposes a GAN-generated image quality objective assessment method AJ-GIQA.Firstly,by simulating the distortion sensitivity of human visual system,it preprocesses the GAN-generated image and obtains the just noticeable difference image.Secondly,by introducing the attention module into the feature extraction network so as to simulate the attention characteristics of human visual system,it obtains the salient features of the images.Finally,the image features are input into the quality prediction network which combines with semantic information,and the quality of the GAN-generated image is comprehensively evaluated based on the image content.The experimental results on the dataset of GAN-generated images show that AJ-GIQA’s assessment results the assessment results of AJ-GIQA are more consistent with the subjective mean opinion scores.The experimental results on the consistency of image quality ranking show that the accuracy of AJ-GIQA is the best in LGIQA-LSUN-cat dataset,and compared with SFA,its other advanced methods,AJ-GIQA’s performance is improved by 0.267.The experimental results on generalization performance show that compared with the most advanced HyperIQA method,the Pearson linear correlation coefficient of AJ-GIQA in dataset PIPAL is increased by 0.027.