雾视错觉理论及应用研究

Visual illusions, as the special phenomena of human visual perception, have been investigated by neuropsychologists, brain scientists and vision researchers with great research enthusiasm for more than a century, in spite of debates concerning both nature and the visual mechanisms involved. Their theories and models have very broad prospects for computer vision and image processing tasks. This research project focuses on the haze illusion, the typical case of lightness illusions, and carries on the empirical research methodology. Firstly, haze illusion systematic experiments are conducted by a preliminary one and a constructive one. Key factors affecting illusory strength are observed serving as the control stimulus strategy with further theoretical analysis. Statistic treatments of modern experimental data are displayed in multi-dimensional form. Then, layered image representations are induced by a novel atmospheric transfer function (ATF) composed of a set of factors through strict dimensionality reduction and coupling. Analyses and inductions of prior assumptions and constraints are summarized to construct the unified solution strategy. Lastly, a common framework for application adaptability of different scenes is built by solution and optimization process of three kinds of new proposed approaches, which are the sophisticated filtering model, variational framework and total variation model. For the typical application scenarios of non-uniform lighting image enhancement, shadow removal and haze removal, image quality improvement is realized sufficiently by these adaptive solutions for ATF models. In such a multi-disciplinary overlapping research area, this project aims at making certain innovative achievements in three levels: theoretical analysis, model solution and practical application. And also, it will lay an enriched foundation for further study and point out the developing direction.

视错觉是人类视觉感知中的特殊现象，百余年来理论解释争鸣不断，但却一直激发着神经心理学、脑科学和计算机科学学者的研究热情，而视错觉理论建模向计算机视觉、图像处理应用转化前景广阔。本项目以亮度错觉典型案例- - 雾视错觉为研究对象，以实证研究为方法论指导，首先系统性开展视觉实验和理论分析：构造两个具体实验场景，探寻新型影响因子，构建控制激励指标集，统计实验数据形成多维表征图；进而实现视觉表征分量获取：强化约减、耦合，综合关键影响因子设计新型环境传输函数，分析归纳模型求解先验假设与约束条件，提出统一求解策略；最终形成场景适应性应用统一框架：提出三种方法(新的滤波估计、变分和全变分框架)进行数学求解和优化，面向三个图像质量提升场景(非均匀光照增强、阴影去除和去雾)进行应用。本项目有望在此交叉研究领域，在理论分析、模型求解和实践应用三个层面取得一定创新性成果，为后续研究强化理论和技术准备、指明途径方向。

本项目以亮度错觉中具有先天优势和特殊性的雾视错觉为重点研究对象，探究人类视觉系统认知本征场景过程中的分层视觉表征内在机理和科学方法，从基础理论、视觉实验和场景应用三个方面在此交叉研究领域进行了一定探索。.1. 深入挖掘雾视错觉基础理论：针对原有雾视错觉中固有问题，重新构建梳理，提出了新型透明耦合环境传输函数，提出了两类X结引导实验设计模式和控制激励指标集，同时分析与透明度感知协同，设计了面向雾视错觉的透明度感知响应类别矩阵。.2. 系统性开展雾视错觉视觉实验：选择合适被试群，依托两种视觉环境分别构造了三个实验场景和环节：视觉适应性、雾视错觉构造性和透明度感知协同实验。记录视觉感觉与匹配过程，数据记录与显示形成一表三图。获得重要结论：(a) 雾视错觉视在感知形成受到乘性和加性两类因素(环境透明度分量和环境亮度分量)影响，在一定范围内环境透明度比值和环境加性分量差值实现引导锚定目标感知，差异值增加导致边界区域亮度比、亮度差比、交叉亮度差比值增大，视在雾感知差异随之增大；(b) 中心结与引导结共同作用生成雾视错觉、透明度感知等视觉现象，透明度感知对雾视错觉生成有帮助作用，无反结具有强化雾视错觉和透明度感知效果，单反结需要区分作用评价区域。.3. 面向多种场景和关键环节应用：从透明层提取、透明度计算、特征生成三个方面对视觉表征分量计算进行了探索、提出了核心方法；在三个场景进行了应用：(a) 面向影像质量提升典型场景:分别提出基于分层视觉表征的单幅图像阴影去除方法、基于新型环境传输函数的单幅图像雾霭去除方法和基于NCL-ICM的真实影像再现框架；(b) 面向目标识别跟踪关键环节：将视觉表征分量计算和特征生成典型方式应用于红外成像制导末端局部图像识别跟踪，提出多种面向无人机视频图像运动目标跟踪方法；(c) 面向空天多源影像视觉解译架构：从空间视频传感器影像分层视觉解译和智能毁伤敏捷评估体系两个方面为应用和架构设计服务，并在视觉解译思维层面提出了一些思考。同时基于理论和算法研究基础，面向无人机影像质量提升与解译等应用，本项目在软硬件系统方面逐步集成，设计了软件系统和嵌入式样机板卡，为后续地面和机载应用奠定了一定基础。