立体图像智能处理关键技术研究
发布时间：2018-07-31 18:01
【摘要】：2009年以来,以阿凡达为代表的3D电影在全球的兴起,带动了立体相机、立体电视等产业市场的高速增长,也推动了多媒体研究领域中立体视觉媒体相关处理技术研究的进步和发展。其中,立体图像处理是对双目立体格式的图像进行加工处理的技术,是支撑和引领相关行业发展的核心技术。充分利用立体图像的双目特性,提高图像处理的质量和效率,将极大推动计算机视觉智能研究的进步和立体视觉消费产业的发展。同传统的图像处理技术相比,立体图像处理技术的关键在于对双目左右视角相关性的挖掘和利用。把传统图像处理技术和视角相关性高效地融合,才能提高立体图像的处理质量。本文针对目前立体图像处理领域的应用需求,从分析立体图像处理的特性入手,对立体图像处理中的深度获取、显著性分析和交互式对象分割三个关键技术展开研究。论文的主要创新和贡献点包括以下几个方面：1.提出了一种新的立体匹配方法,通过结合视角相关性与颜色相似性,有效地解决了现有方法过度依赖颜色分布而导致的精度不足和鲁棒性较差等问题。现有的局部立体匹配方法主要基于单个视角内容的颜色相似性。这类方法过度依赖于图像颜色和深度分布的一致性,在实际应用中的精度和鲁棒性都不高。本文通过分析立体图像左右视角的相关性,并与单视角的颜色相似性线索相结合,提出了一种将两者进行互补的代价聚集方法。该方法不仅提高了立体匹配的精度,在实际应用中也表现出更好的鲁棒性。2.提出了一种基于立体视觉的显著性分析方法,其检测结果要明显优于现有最好的方法。 目前的视觉显著性分析主要以二维平面图像作为视觉输入,未能有效地挖掘场景深度信息的潜力,并且缺少相关的数据集支撑基于深度的显著性研究。本文以立体图像作为视觉输入,重点讨论了利用立体图像中隐藏的深度信息进行显著对象检测的问题。通过考察深度图像和显著对象空间结构的特性,本文提出了一种基于深度各向异性对比度的显著区域检测方法,并与现有的基于二维图像的方法进行了结合。在实验评测中,本文方法的精度和F值都比现有最好方法高出15%以上。此外,针对目前缺少基于深度的显著对象检测数据集的现状,本文公布了目前国际上最大的基于深度的显著对象检测数据集,以推动本领域的相关研究。3.提出了一种服务于立体图像的一致性分割方法,大幅提升了立体图像左右视角联合分割的处理效率。现有的视角一致性分割方法大多是基于区域的,没有考虑到立体图像视角内容间的相似性,导致计算冗余度过高。针对该问题,本文提出了一种基于轮廓的一致性分割方法,其计算效率达到现有最快方法的10倍,并且具有更好的一致性分割精度。此外,本文方法采用了对一致性约束独立求解的思路,因而可以与任何现有的单视角分割方法无缝结合,共同处理立体图像分割问题。4.提出了一种基于颜色和深度自适应融合的对象分割方法,显著地提升了对象分割的精度。 目前的图像分割方法大多是基于颜色信息的,未能对深度信息的特性进行充分的挖掘。本文通过考察深度图像的特点,提出了一种基于测地距离的对象分类模型,并进一步通过分析深度与颜色信息在图像分割中各自的特性,对颜色与深度信息进行了自适应融合。该方法在两个相关评测平台上比现有最好方法分别取得了2%和3%的F值提升,并且精度和召回率指标都高达95%以上。在以上关键技术研究的基础上,本文还给出了一系列立体图像智能处理的案例,从而展示了本文研究成果在相关应用领域中具有很好的支撑作用和应用前景。
[Abstract]:Since 2009, the rise of 3D films represented by arfan has led to the rapid growth of stereoscopic cameras, stereoscopic television and other industrial markets, which also promoted the progress and development of the research on stereoscopic visual media related processing technology in the field of multimedia research. The processing technology is the core technology to support and lead the development of the related industries. Making full use of the binocular features of stereoscopic images, improving the quality and efficiency of image processing, will greatly promote the progress of computer vision intelligence research and the development of stereoscopic visual consumption industry. The key lies in the mining and utilization of the relativity between the eyes and the eyes. In order to improve the quality of the stereo image processing, the quality of the stereo image processing can be improved by combining the traditional image processing technology with the angle of view, so this paper, aiming at the application requirement of the stereo image processing field, starts with the analysis of the characteristics of the stereo image processing, and the depth acquisition in the stereo image processing. The main innovations and contributions of this paper are as follows: (1.) a new method of stereo matching is proposed, which can effectively solve the accuracy of the over dependent color distribution of the existing method by combining the angle of view with the color similarity. The existing local stereo matching methods are mainly based on the color similarity of the single view content. This method is overly dependent on the consistency of the color and depth distribution of the image, and the accuracy and robustness in the practical application are not high. This paper analyzes the correlation of the angle of view of the stereo images and is with the single angle of view. With the combination of color similarity cues, a method of cost aggregation is proposed to complement each other. This method not only improves the accuracy of stereo matching, but also shows a better robustness in practical application..2. proposes a method of saliency analysis based on stereoscopic vision. Its detection results are obviously superior to the best existing methods. The current visual saliency analysis mainly uses two-dimensional plane images as visual input, fails to effectively excavate the potential of depth information of the scene, and lacks the related data sets to support the significance of depth based research. This paper uses stereoscopic images as visual input, focusing on the use of hidden depth information in stereoscopic images. In this paper, a significant region detection method based on the depth anisotropy contrast is proposed and combined with the existing two dimensional image based method. In the experimental evaluation, the accuracy and F value of this method are better than the existing best methods. In addition, in view of the current lack of a significant object detection data set based on depth, this paper publishes the largest set of significant object detection data based on depth in the world, in order to promote the related research in this field,.3. proposed a method of conformance segmentation for stereoscopic images and greatly enhanced the stereoscopic image. Most of the existing angle of view conformance segmentation methods are based on the region, which does not take into account the similarity between the content of the stereoscopic image, which leads to high computational redundancy. In this paper, a contours based conformance segmentation method is proposed in this paper, and its computational efficiency reaches the fastest method available. 10 times, and has better conformance segmentation precision. In addition, this method adopts the idea of independent solution to the consistency constraint, so it can be combined with any existing single view segmentation method to deal with the problem of stereo image segmentation..4. proposes an object segmentation method based on color color and depth adaptive fusion. Most of the current image segmentation methods are based on the color information and fail to fully excavate the characteristics of the depth information. In this paper, an object classification model based on the distance of the geodesic is proposed by investigating the characteristics of the depth image, and the depth and color information are analyzed in the image. The characteristics of the segmentation are adaptive fusion of color and depth information. The method has obtained 2% and 3% F values on the two related evaluation platform, and the precision and recall index are up to 95%. Based on the key technology research above, a series of stereoscopic images are given. Intelligent processing cases show that the research results of this paper have a good supporting role and application prospects in the related fields.
【学位授予单位】：南京大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41
[Abstract]:Since 2009, the rise of 3D films represented by arfan has led to the rapid growth of stereoscopic cameras, stereoscopic television and other industrial markets, which also promoted the progress and development of the research on stereoscopic visual media related processing technology in the field of multimedia research. The processing technology is the core technology to support and lead the development of the related industries. Making full use of the binocular features of stereoscopic images, improving the quality and efficiency of image processing, will greatly promote the progress of computer vision intelligence research and the development of stereoscopic visual consumption industry. The key lies in the mining and utilization of the relativity between the eyes and the eyes. In order to improve the quality of the stereo image processing, the quality of the stereo image processing can be improved by combining the traditional image processing technology with the angle of view, so this paper, aiming at the application requirement of the stereo image processing field, starts with the analysis of the characteristics of the stereo image processing, and the depth acquisition in the stereo image processing. The main innovations and contributions of this paper are as follows: (1.) a new method of stereo matching is proposed, which can effectively solve the accuracy of the over dependent color distribution of the existing method by combining the angle of view with the color similarity. The existing local stereo matching methods are mainly based on the color similarity of the single view content. This method is overly dependent on the consistency of the color and depth distribution of the image, and the accuracy and robustness in the practical application are not high. This paper analyzes the correlation of the angle of view of the stereo images and is with the single angle of view. With the combination of color similarity cues, a method of cost aggregation is proposed to complement each other. This method not only improves the accuracy of stereo matching, but also shows a better robustness in practical application..2. proposes a method of saliency analysis based on stereoscopic vision. Its detection results are obviously superior to the best existing methods. The current visual saliency analysis mainly uses two-dimensional plane images as visual input, fails to effectively excavate the potential of depth information of the scene, and lacks the related data sets to support the significance of depth based research. This paper uses stereoscopic images as visual input, focusing on the use of hidden depth information in stereoscopic images. In this paper, a significant region detection method based on the depth anisotropy contrast is proposed and combined with the existing two dimensional image based method. In the experimental evaluation, the accuracy and F value of this method are better than the existing best methods. In addition, in view of the current lack of a significant object detection data set based on depth, this paper publishes the largest set of significant object detection data based on depth in the world, in order to promote the related research in this field,.3. proposed a method of conformance segmentation for stereoscopic images and greatly enhanced the stereoscopic image. Most of the existing angle of view conformance segmentation methods are based on the region, which does not take into account the similarity between the content of the stereoscopic image, which leads to high computational redundancy. In this paper, a contours based conformance segmentation method is proposed in this paper, and its computational efficiency reaches the fastest method available. 10 times, and has better conformance segmentation precision. In addition, this method adopts the idea of independent solution to the consistency constraint, so it can be combined with any existing single view segmentation method to deal with the problem of stereo image segmentation..4. proposes an object segmentation method based on color color and depth adaptive fusion. Most of the current image segmentation methods are based on the color information and fail to fully excavate the characteristics of the depth information. In this paper, an object classification model based on the distance of the geodesic is proposed by investigating the characteristics of the depth image, and the depth and color information are analyzed in the image. The characteristics of the segmentation are adaptive fusion of color and depth information. The method has obtained 2% and 3% F values on the two related evaluation platform, and the precision and recall index are up to 95%. Based on the key technology research above, a series of stereoscopic images are given. Intelligent processing cases show that the research results of this paper have a good supporting role and application prospects in the related fields.
【学位授予单位】：南京大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41
【相似文献】
相关期刊论文 前10条
1 杨s，
本文编号：2156391
本文链接：https://www.wllwen.com/shoufeilunwen/xxkjbs/2156391.html
