基于多变量自然场景统计和局部均值估计的无参考立体图像质量评价

A research on building a computational model for no-reference stereoscopic image quality assessment (NR-SIQA) will be conducted from two key aspects of the image feature extraction and the subjective quality mapping. Since joint distribution among scales or orientations of disparity information can stably characterize the quality of stereoscopic images, which can not be modeled by the univariate nature scene statistics-based approach, the multivariate natural scene statistics models are utilized to extract binocular disparity features. With the support of the stereoscopic visual saliency computational model, the obtained binocular disparity features will be combined with binocular fatigue features to build 3D image quality descriptor. In addition, some existing 2D image quality descriptors will be selected, combined and improved to work with the proposed 3D image quality descriptor to obtain stable and robust quality-aware features; On the other hand, the novel subjective quality mapping algorithm based on local means will be explored to resolve the problem of the traditional global mapping approaches. A classifier that checks whether two images are similar in terms of image quality is employed to determine some images with calibrated subjective scores from image quality database whose feature vectors are similar to the image to be evaluated, and then the quality score of the image is estimated by local means method. This research project will enrich modeling techniques of the natural scene statistics and expand its application scope, and the obtained subjective quality mapping algorithm has better performance in terms of complexity, robustness, scalability and accuracy than the existing algorithms.

从图像质量特征提取和图像质量值映射两个关键方面就构建无参考立体图像质量评价计算模型的相关问题展开研究。针对视差信息在各尺度间和各方向间的联合分布能稳定反映3D图像质量优劣的特点而单变量自然场景统计模型无法对此建模的问题，研究多变量自然场景统计模型以提取双目视差特征，同时结合双目疲劳度特征并在立体视觉显著度计算模型支撑下构建3D质量感知描述符。筛选、组合和改进现有的2D质量感知描述符与3D质量感知描述符联合使用，从而获得完备和鲁棒的立体图像质量感知描述符；探索基于局部均值估计技术的新型图像质量值映射算法，用图像质量同等分类器先大致在图像质量数据库中圈定若干与待评价图像特征矢量近似的图像(已标定质量值), 然后基于这些图像用局部均值估算出图像质量预测值。项目的研究将丰富和拓展自然场景统计建模理论及其应用范围，新型映射算法复杂度低，有望在鲁棒性、可扩展性、准确率等方面较现有算法有更好的性能。

立体图像质量客观评价(stereoscopic image quality assessment, SIQA)主要用于评价立体图像质量的优劣，可作为各种立体图像处理系统参数设置和性能优化依据，具有重要的理论研究和广泛的应用价值。SIQA算法终极的设计目标是实现能自动地获取与人眼感知立体图像质量好坏相一致的计算模型。由于目前人类视觉系统复杂的工作原理并未被完全掌握，设计出与人类视觉系统评价结果相一致的计算模型具有极大挑战。当前大多数的SIQA算法研究仅探索了视觉系统的部分功能与特性，仍存在着对双目感知特性考虑不充分、对非对称失真立体图像预测准确性不高、特征提取时间过长等问题。为此，本项目进一步探究人类双目视觉的感知特性，从图像质量特征提取和图像质量值映射两个关键方面就构建SIQA计算模型的相关问题展开研究。具体地，筛选、组合和改进现有的2D质量感知描述符与3D质量感知描述符联合使用，利用两个视点图像的梯度幅值响应模拟双眼视觉输入刺激，基于眼优势原理将左、右视点分别作为主视图来合成两幅不同的融合图像，用于提取双目感知特征，从而获得完备和鲁棒的立体图像质量感知描述符。在图像质量值映射方面，探索了基于局部均值估计和自适应增强支持向量回归预测模型，实现了更好的感知特征向量到图像质量的映射。在包含了左、右视点失真类型不同的非对称失真图像的Waterloo IVC 3D Phase I(WIVC-I)和Waterloo IVC 3D Phase II(WIVC-II)基准测试数据库上，所研究SIQA算法的SROCC和PLCC相关系数指标值均在0.95以上，明显高于(优于)其他对比算法，而RMSE(root mean square error)指标值分别为3.8623和4.6591，远低于(优于)其他对比算法。本项目所建立的SIQA算法符合人类视觉特性，其研究丰富和拓展了自然场景统计建模理论及其在图像质量评价领域的应用范围。

内容获取失败，请点击重试