基于CNN-SVR的立体图像视觉感知评价方法  

Visual Perception Evaluation Method of Stereo Images Based on CNN-SVR

作　　者：徐源 陈纯毅[1] 胡小娟[1] 于海洋[1] 田野[1] Xu Yuan;Chen Chunyi;Hu Xiaojuan;Yu Haiyang;Tian Ye(School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,Jilin,China)

机构地区：[1]长春理工大学计算机科学技术学院,吉林长春130022

出　　处：《激光与光电子学进展》2023年第8期339-347,共9页Laser & Optoelectronics Progress

基　　金：国家自然科学基金项目(U19A2063);吉林省科技发展计划项目(20230201080GX)。

摘　　要：针对立体图像的多维影响因素和预测结果准确性不足的问题,提出一种基于卷积神经网络-支持向量回归(CNNSVR)的立体图像视觉感知客观评价模型。该模型将基于颜色的平面显著图和基于差异的视差图相结合,对其进行阈值分割,得到视觉感知潜在显著不适区域;然后进行特征提取,分别提取对比度、颜色、结构复杂度等全局特征和视差、纹理、空间频率等局部特征;最后采用将CNN和SVR相结合的方式构建多特征视觉感知客观评价模型,得到最终的客观预测值。实验结果表明,所提方法的Pearson相关系数高于0.87,Spearman相关系数高于0.83。与现有其他方法相比,在公开数据集上所提客观评价模型更优,预测结果与人们主观评价结果具有更高的一致性。This paper proposes an objective evaluation model of stereo image visual perception based on convolutional neural network(CNN) and support vector regression(SVR) to solve the issue of multidimensional influencing factors for stereo images and improve the accuracy of prediction results.In this proposed model,the plane saliency map using color and the disparity map based on differences are combined,and the potential salient discomfort regions of visual perception are obtained using threshold segmentation.Then,global features,such as contrast,color,and structural complexity,are extracted using feature extraction along with the local features,such as disparity,texture,and spatial frequency.Finally,the objective evaluation model of multifeature visual perception is constructed by combining CNN and SVR to obtain the final objective prediction value.Experimental results show that the Pearson linear correlation coefficient and Spearman's rank correlation coefficient of the proposed method are higher than 0.87 and 0.83,respectively.In addition,compared with other existing methods,the objective evaluation model proposed in this paper is better on the public dataset,and the prediction results have higher consistency with the subjective evaluation results.

关 键 词：立体图像 视觉感知 特征提取 卷积神经网络 支持向量回归 客观评价模型 

分 类 号：TN911.73[电子电信—通信与信息系统]