基于图像非平坦区域DCT特性和EGRNN的盲图像质量评价

摘要

对图像的质量进行评价有着广泛的应用,为定量的衡量图像的降质水平,文中在无指定图像失真类型的情况下,提出了一种通用的盲图像质量评价(Blind Image Quality Assessment,BIQA)方法.该方法首先提出了图像平坦区域和非平坦区域及视觉屏蔽的概念.将分割图像的平坦和非平坦区域引入评价系统中,从图像的非平坦区域中提取出能反应图像质量的视觉特征(Visual Features,VF),视觉特征涵盖了图像空域特性和DCT变化后的频域特性.利用非平坦区域中提取的空域与频域的联合VF和差分平均意见得分(Difference Mean Opinion Score,DMOS)训练提出的均衡广义回归神经网络(Equalization General Regression Neural Network,EGRNN).最后用LIVE的IQA数据库Release2测试了该方法.实验表明,分割图像非平坦区域,降低了平坦背景对图像质量评价的干扰,提升了预测的稳定性和准确性.此外,EGRNN通过均衡层的引入解决了视觉屏蔽现象,对图像质量的预测效果高于GRNN (General Regression Neural Network).该BIQA方法较其他方法预测更准确、更稳定.%The image quality assessment applied very extensively.To quantify image degradation level,a distortion-agnostic,general-purpose BIQA (Blind Image Quality Assessment) method is proposed in this paper.Firstly,the concepts of image flat region,non-flat region and visual mask are proposed.Secondly,the segmentation of the image flat region from non-flat region are introduced to the measurement system.Thirdly,the visual features are extracted from the non-flat region of images,which include the spatial domain and frequency domain features.Those features are shown to correlate highly with image quality.Fourthly,towards ameliorating the predicted model against the impact of the visual mask,the present GRNN (General Regression Neural Network) is improved with the tradeoff matrix.Then,an advanced predicted model,called EGRNN(Equalization General Regression Neural Network),is trained by the VF (Visual Features) and the DMOS (Difference Mean Opinion Score).Finally,the predict performance of proposed algorithm is tested on Database Release2 from LIVE.The experimental result demonstrates that the non-flat region segmentation reduces the interference of monochromatic background,improves the prediction accuracy of image quality.Besides,avoided the visual mask,EGRNN shows a better predict performance than GRNN.The predict performance of proposed method is higher than the other approaches on accuracy,monotonicity and reliability.