基于多域图像转换的多等级真实失真图像生成算法

A multi-level authentic distortion image generation algorithm based on multi-domain image-image transition

为进一步提升真实失真下图像质量评估算法的准确性,提出一种基于多域图像转换的多等级真实失真图像生成算法。利用已有的真实失真数据集训练基于生成对抗网络的图像转换模型,使模型能够将高清图转换为不同等级的失真图,且视觉上近似真实失真。按照失真等级对图像进行标签标注,实现用标签控制模型生成多个等级的失真图像。通过实验数据证明了所提算法能够生成失真等级分明的图像,且相比于合成失真算法更接近真实情况。峰值信噪比指标下测试集的测试准确率为76%,且模型参数量少于同类型模型。对比实验表明,使用生成的不同等级图像预训练,可以帮助模型在评估中取得更好的评估结果。

To further improve the accuracy of image quality assessment algorithm under true distortion,a multi-level authentic distortion image generation algorithm based on multi domain image conversion was proposed.The existing real distortion data set was used to train the image conversion model based on the generation countermeasure network,so that the model could convert the high-definition image into different levels of distortion map,and visually approximate the true distortion.Image was labeled according to the distortion level,and multiple levels of distorted image were generated with label control model.Experimental data show that the proposed algorithm can generate images with distinct distortion levels,and it is closer to the real situation than the synthetic distortion algorithm.Under the index of peak signal to noise ratio,the accuracy of the test set is 76%,and the parameters of the model are less than those of the same type.Comparative experiments show that different levels of image pre-training can help the model to achieve better evaluation results in the evaluation.