基于多尺度特征的无参考屏幕内容图像质量评估

No Reference Image Quality Assessment for Screen Content Image Based on Multi-scale Features

屏幕内容图像是一种组合图像,由计算机将图形、文字和图像组合起来而形成.由于人类视觉系统是从粗略到精细进行图像信息的提取,本文提出一种基于多尺度特征的无参考屏幕内容图像质量评估算法.屏幕内容图像中包含大量的图形和文本内容,以及色彩和布局结构信息,因此我们提取失真图像的边缘特征、结构特征和亮度特征.然后将多个图像尺度上提取的特征进行拼接,作为最终的失真图像质量感知特征.最后使用随机森林回归方法训练得到无参考屏幕内容图像质量评估模型.实验结果表明,本文引入的多特征和多尺度机制是有效的,相比较目前先进的无参考方法,本文模型可以取得与主观感知更高的一致性,甚至在整体性能上超过了多数全参考方法.

The screen content image is a kind of combined image,which is formed by a computer combining graphics,text and images.Since the human visual system extracts image information from rough to fine,this paper proposes ano-reference screen content image quality evaluation algorithm based on multi-scale features.The screen content image contains a large amount of graphical and textual content,as well as color and layout structure information,so we extract the edge features,structural features,and color features of the distorted image.Then the features extracted from image with multiple scalesare stitched together as the final quality perception feature of distorted image.Finally,the random forest regression method is used to train the no-reference screen content image quality evaluation model.The experimental results show that the multi-feature and multi-scale mechanisms introduced in this paper are effective.Compared with the current advanced no-reference methods,the model in this paper can achieve higher consistency with subjective perception,and even surpass most full-reference methods in overall performance.