摘要： 针对支持向量回归(SVR)中惩罚因子和径向基函数选取具有较大不确定性和随机性的问题,结合单双目信息与基于遗传算法(GA)的SVR优化模型,提出无参考立体图像质量评价方法。提取左右失真图像的单双目特征,将梯度幅值和拉普拉斯特征作为单目视觉特征。为更好地结合人类双目视觉特性,使左右图像融合成一幅独眼图,对独眼图提取空域自然场景统计特征。利用GA选择、交叉和变异等操作优化SVR参数组合,选出最优的参数组合,引入到SVR中预估左右图像质量。考虑到人眼对于左右失真图像的响应不同,通过增益控制模型融合左右图像质量,从而得到最终的质量评价值。应用该评价方法对宁波大学建立的立体数据测试库进行评价,结果表明其Pearson线性相关系数在0.95以上,Spearman等级相关系数值在0.94以上,与人类主观感知具有高度一致性。

关键词: 遗传算法, 支持向量回归, 立体图像质量评价, 单双目视觉特性, 增益控制模型

Abstract: Aiming at the problem of the uncertainty and randomness of the selection of penalty factor and radial basis function in the Support Vector Regression(SVR),based on the monocular and binocular information as well as the SVR optimization model based on Genetic Algorithm(GA),a non-reference stereoscopic image quality assessment method is proposed.The monocular and binocular visual features of the stereoscopic images are extracted.The gradient magnitude and laplacian features are used as monocular visual features.Then,in order to better combine the binocular visual features of human beings,left and right distorted images are fused to construct a cyclopean map from which the statistical characteristics of natural scenes in the spatial domain are extracted.The SVR parameter combination is optimized by using selection,crossover and mutation operations of the GA.The optimum parameter composition is introduced into SVR to predict the objective scores of the stereoscopic images.Considering the responses of human eyes to distortions of the left and right images are different,the gain control model is adopted to fuse the left and right image quality to get final quality evaluation value.Experimental results show that The Pearson Linear Correlation Coefficient(PLCC) indicators reach 0.95 and Spearman Rank-order Correlation Coefficient(SROCC) indicators reach 0.94 on stereoscopic image database established by Ningbo University.The proposed algorithm is highly consistent with the subjective perception.

Key words: Genetic algorithm(GA), Support Vector Regression(SVR), stereoscopic image quality assessment, monocular and binocular visual features, gain control model

中图分类号: 

• TN911.73