基于对抗和梯度的无参考图像质量评价算法

No⁃reference image quality assessment algorithm based on adversarial and gradient

由于现有的生成对抗网络(GAN)很难完全获取细节上的失真,生成高质量图像仍旧比较困难。为了提高基于GAN的无参考图像质量评价(NR⁃IQA)方法的性能,提出一种基于对抗和梯度的NR⁃IQA方法。通过改进网络模型结构和引入基于梯度和色度相似性的平均偏差相似指数(MDSI)来提升模型的整体性能,整个模型由GAN和质量预测网络组成。首先,为了增强对抗训练,设计了双判别器的GAN,将失真图像和参考图像分别输入到网络中,利用WED数据集对GAN进行训练;其次,利用GAN生成相应失真图像的虚拟图像,并分析两者之间的差异,得到失真差异图和MDSI图;最后,为了从多个方面测量图像的感知质量,设计了多流质量预测网络,将失真图像、虚拟图像、失真差异图、MDSI图分别输入到网络中,输出图像质量预测分数。在LIVE、CISQ、TID2013数据集上进行训练和测试,所提算法在三个数据集上都表现出较好的性能,尤其在TID2013上。实验结果表明,该算法与人的主观评价具有较高的一致性。

Generating high quality images still remain difficult due to the difficulty of fully capturing distortion in detail with existing generative adversarial networks(GAN).To improve the performance of the GAN⁃based no⁃reference image quality assessment(NR⁃IQA)method,an NR⁃IQA method based on adversarial and gradient is proposed.The overall performance of the model is improved by improving the network model structure and introducing the mean deviation similarity index(MDSI)based on gradient and chromaticity similarity,and the whole model is consisted of GAN and quality prediction network.In order to enhance the adversarial training,a dual discriminator GAN is designed,where the distorted image and the reference image are input to the network separately,and the GAN is trained by means of the WED dataset.The GAN is used to generate virtual images of the corresponding distorted,and the differences between them are analyzed to obtain distortion difference maps and MDSI maps.In order to measure the perceptual quality of the images from multiple aspects,a multi⁃stream quality prediction network is designed,and the distorted image,virtual image,distortion difference map and MDSI map are input to the network separately to output the image quality prediction score.The proposed algorithm was trained and tested on the LIVE,CISQ and TID2013 datasets.This algorithm has shown good performance on all three datasets,especially on the TID2013.The experimental results show that this algorithm has high consistency with human subjective evaluations.