基于近邻算法的GAN生成图像质量评价

GAN generated image quality assessment based on the near-neighbor algorithm

GAN生成图像质量评价是指对GAN生成的图像进行评价,判断生成图像的失真度是否影响观察者的信息获取和主观感受.目前,GAN生成图像质量评价算法较少且算法运行效率不高.该文提出一种基于近邻算法的生成图像质量评价(Near-Neighbor based Generated Image Quality Assessment,NN-GIQA)算法,实现对GAN生成图像的自动、客观、高效评价.首先,基于ANN算法获取生成图像的近邻构成相似图像候选池,缩小生成图像对比范围;然后,基于KNN算法在相似图像候选池中获取与生成图像最相似的K个真实图像得到生成图像质量分数;最后,评价多个经典GAN模型在多个经典数据集上获取的生成图像的质量.实验结果表明本文方法有效提高了GAN生成图像质量评价的效率和准确性,运行时间仅为其他方法的1/9~1/28,其评价结果和人类主观评价结果的一致性达到80%以上,符合人类视觉感知.

GAN-generated image quality assessment is to assess the image generated by GAN and judge whether the distortions of the generated image affect the observer’s information acquisition and subjective perception.Currently,the assessment methods for GAN generated image quality are few and have low operation efficiency.This paper proposes a generated image quality assessment method based on the near-neighbor algorithm(Near-Neighbor based Generated Image Quality Assessment,NN-GIQA)to assess the image generated by GAN automatically,objectively,and efficiently.Firstly,the near neighbors of the generated images were acquired based on the ANN algorithm to form a similar image candidate pool and reduce the range of generated images for comparison.Secondly,based on the KNN algorithm the K real images most similar to the generated image from the similar image candidate pool are obtained,and the generated image quality score is given.Finally,assess the generated image quality obtained by multiple classical GAN models on multiple classical datasets.The experimental results show that the method effectively improves the efficiency and accuracy of GAN-generated image quality assessment,the running time is only 1/9~1/28 of other methods,and the consistency between the assessment results and human subjective assessment results reaches more than 80%,assessment results consistent with human visual perception.