SIA OpenIR  > 机器人学研究室
图像光照建模与处理算法研究
其他题名Image Illumination Modeling and Processing
屈靓琼1,2
导师唐延东
分类号TP391.41
关键词图像光照建模 图像光照处理 深度学习 阴影去除 阴影特征
索取号TP391.41/Q87/2017
页数105页
学位专业模式识别与智能系统
学位名称博士
2017-11-30
学位授予单位中国科学院沈阳自动化研究所
学位授予地点沈阳
作者部门机器人学研究室
摘要复杂多变的光照环境是制约相机成像质量的主要因素之一,给计算机视觉算法的应用与发展(如物体识别与跟踪、场景理解)带来诸多不利影响,降低了其算法的鲁棒性及环境自适应性。该问题的解决,将会促进计算机视觉和相关学科的发展,具有广阔的应用前景。不同于目前主流的基于机器学习或基于数学的方法中单纯采用数据驱动的研究方式,本文从物理成像的基本原理和深度学习两个方面开展了图像光照建模与处理算法研究,主要研究内容包括光照建模、光照不变图像、阴影去除及光照处理算法在计算机视觉中的应用。论文中主要创新点如下: (1) 以局部光照变化阴影为研究对象,对几种常用的阴影特征进行了综合考核和评估。该项评估表明,有效的阴影特征是影响图像阴影识别、去除及恢复等算法性能的关键因素。通过特征分析和实验对比, 我们发现,目前数据驱动的阴影特征具有二义性或多义性,无法有效表征阴影的独有特性。基于此,我们对目前通用的图像阴影特征,给出了性能排序、局限性及其有效应用的场景范围。这是目前图像处理领域首次开展的评估阴影特征的工作,为后续的光照建模和阴影处理算法等问题的解决提供了重要的指导意义。(2) 针对室外光照环境,以现有的线性阴影模型为研究基础,为图像的每个像素建立一个线性方程组,并对该线性方程组的解进行正交分解,建立了具有原创性的图像光照像素级正交分解模型。不同于目前主流的纯数据驱动的光照处理算法,该模型将物理成像机理、图像数据及数学线性代数理论相结合,具有合理的物理解释和严格的数学推导。该数学推理与物理成像机理结合的方式为突破现有方法歧义性强、实时性差和缺乏物理解释的瓶颈问题,提供了一种有效的理论模型。(3) 针对室内室外两种光照环境,以局部光照变化阴影为研究对象,结合阴影形成机理和深度学习技术,建立了一种基于深度学习的光照模型。该模型将阴影检测以多上下文神经网络的方式隐式嵌入到阴影去除中,解决了现有阴影去除算法过渡依赖阴影检测和人工特征的问题。该模型首次实现物理成像机理与深度学习技术的融合,无需依赖先验知识和经验假设,稳定性强、普适性好,且不受光照环境和天气影响,为光照处理领域的其他问题及底层图像处理算法提供了新的研究思路。(4) 在物理光照模型基础上,提出了具有光照自适应能力的物体颜色恒常算法。针对目前物体颜色恒常算法依赖空间域一致性假设或者光照变化平缓假设,以本文物理光照模型基础,计算场景内光照变化比率和转移向量,建立了基于正交分解模型的多光源下的物体颜色恒常算法;该算法不依赖经验假设也无需复杂的特征检测等,实时性强、可移植性高,能直接应用到后续的物体识别和跟踪算法等任务上,提高相关算法的性能。(5) 在深度光照模型的基础上,提出了具有光照自适应能力的RGBD图像显著性区域检测算法。该算法借助深度学习的强大学习能力,将低层光照不变的图像显著特征组合形成为更加抽象的高层表示,结合深度光照模型过滤误判区域,解决了底层特征表示受光照变化影响的局限性及各特征之间相互作用机制不明朗等问题。理论分析和实验结果均验证了该算法的有效性和先进性。
其他摘要As a main factor of imaging, the complex illumination variation often causes a lot of problems for a variety of computer vision tasks and its application, such as object recognition and tracking and scene understanding. It degrades the performance of the algorithms in computer vision tasks, and impedes the adaptiveness of the algorithm to different and complex illuminant conditions. Image illumination modelling and processing are of great practical significance and have attracted a lot of attentions in recent years. Different from current works that purely analyze on the image data with mathematics theories or machine learning techniques, in this dissertation, we study the image illumination modeling and processing problems based on physically imaging mechanism and deep learning techniques. We investigate a number of illumination related issues, including image illumination modeling, illuminant invariant images, shadow removal algorithms, and the applications of these illuminant algorithms in several computer vision tasks. The main research contributions of this dissertation are summarized as follows: (1) From the view of local illumination variation, a comprehensive evaluation of image shadow features, which are commonly applied in image shadow processing, are presented. This evaluation indicates that shadow feature is the key factor that influence the performance of shadow detection and shadow removal. The feature analyses and experiment comparisons show that these purely image data based shadow features are often ambiguous and can not characterize the specificity of shadow regions effectively. In the evaluation, the performance ranking, the limitations and the effectiveness of different shadow features are presented. To the best of our knowledge, this is the first work to evaluate shadow features, which can offer guidance for future illumination modeling and shadow processing algorithms. (2) For outdoor illumination, a novel and effective pixel-wise orthogonal decomposition for color illumination invariant is proposed. In this model, for each RGB pixel value, a linear equation set is set up based on the existing shadow linear model. Through orthogonal decomposition of the solution space of this linear equation set, a color image can be directly decomposed in pixel-wise into an illumination invariant image and the illuminant intensity. Unlike traditional image data based illumination processing methods, this model combines physical image mechanism, image data and mathematical linear algebraic theory, and has a reasonable physical interpretation and rigorous mathematical deduction. This combination mechanism provides new insight for the current research, and solves the problems of ambiguity, poor real-time, and lack of physically meaning in the current image illumination modeling and processing methods. (3) To handle the local illumination variation in both outdoor and indoor environment, a universal deep learning based illumination model is proposed by combining the shadow formulation mechanism with deep learning techniques. Unlike the conventional methods that require shadow detection or only utilize hand-crafted features, a multi-context embedding deep network is designed to learn the mapping function between a shadow image and its illumination attenuation. This model is the first work that tries to combine physically imaging mechanism with deep learning techniques to handle local illumination variation. It does not impose any assumptions on the light sources nor require any empirical hypothesis. Thus, it is adaptive to different and complex illuminant conditions. This kind of multi-context embedding deep network provides new insight for the research on low-level computer vision tasks, and can be applied to handle other complex illumination variations tasks. (4) Based on the physical illumination model, an illumination invariant object color constancy method is proposed. Unlike conventional color constancy methods that impose spatial uniformity assumption or smooth illuminant transaction assumption on the scene, in this method, a transfer vector and the illuminants ratio vector are first calculated based on physical illumination model, and then an object color constant image for outdoor multiple light sources is obtained by a simple pixel-wise orthogonal decomposition operation. Without any empirical assumptions or any feature based detection, this method can be directly applied to some real-time applications, such as object recognition and tracking, and can improve the performance of these methods. (5) Based on the deep illumination model, an illumination invariant RGBD salient object detection method is proposed. This method adopts the deep learning techniques to automatically combine several low level illumination invariant saliency feature vectors into unified and representative hyper features, and then combines the deep illumination model to filter out the wrongly detected shadow regions. This proposed method solves the problems of conventional methods that are vulnerable to illumination variations and confined by the uncertainty of the interaction mechanisms between different low level saliency features. Both the theory analyses and comparison experiments demonstrate the effectiveness and powerfulness of this proposed method.
语种中文
产权排序1
文献类型学位论文
条目标识符http://ir.sia.cn/handle/173321/21277
专题机器人学研究室
作者单位1.中国科学院沈阳自动化研究所
2.中国科学院大学
推荐引用方式
GB/T 7714
屈靓琼. 图像光照建模与处理算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
图像光照建模与处理算法研究.pdf(70307KB)学位论文 开放获取CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[屈靓琼]的文章
百度学术
百度学术中相似的文章
[屈靓琼]的文章
必应学术
必应学术中相似的文章
[屈靓琼]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。