|Thesis Advisor||张晓鹏 ; 严冬明|
|Place of Conferral||智能化大厦3层|
|Keyword||保边滤波器加速 保边滤波器设计 图像去噪 网格去噪 可视数据集合总结|
Visual data is the type of data that carries visual information, including images, videos, and geometric models. With the development and popularization of visual data acquisition equipments, such as digital cameras and 3D scanners, users can obtain images, videos, 3D models and other visual data more easily. Advances in Internet technology have also made more visual data accessible. The vivid and intuitive visual data provides sufficient information for users, which will enrich their lives and improve the work efficiency. At the same time, effective acquisition, efficient computing and content analysis have become the key issues in visual data computing. First, the acquisition and transmission of visual data often introduce noises, which degrade the data quality. Second, the resolution of visual data, such as images, videos and 3D meshes, is constantly improved, which prevent users from processing visual data in real time. Finally, from the overall perspective of the visual data collection, the lack of organization of data elements and information redundancy bring difficulties for users to understand the content of the visual data collection and to retrieve valuable information. To infer the most desired information from various visual data, it is highly-demanded for effective and efficient visual data processing and analysis technologies.
For individual visual data elements, edge-preserving smoothing can separate edges and sharp structures in images and meshes from noises and details. Designing fast edge-preserving filter with excellent edge-preserving ability can help realizing high-quality and efficient visual data processing. For the whole visual data collection, it will be convenient for users to obtain desired information and understand the visual data collection by well organizing the elements and displaying the diverse, preferred, and representative elements as the summarization of the collection. To process and analyze visual data effectively and efficiently, three problems are studied in this thesis: accelerating and designing edge-preserving filters for images and meshes, and summarizing the visual data collection according to user preference. The main contributions are as follows: (1)An accurate bilateral filter acceleration method using weighted variable projection to achieve linear time complexity is proposed. The previous bilateral acceleration methods had poor accuracy in bilateral filtering results approximation. In the thesis, the range kernel approximation problem is formulated into nonlinear joint optimization about the basis functions and the corresponding coefficients. Meanwhile, the information of image color distribution is also taken into account in the optimization objective. The weighted variable projection technique is utilized to solve this problem to obtain high precision approximation results and linear time acceleration algorithm. Experiments demonstrate that the proposed weighted variable projection method can obtain accurate range approximation results, and can gain more accurate filtering results efficiently.
(2) A space-variant bilateral filter with better edge-preserving ability and its error-bounded linear time acceleration method is proposed. The traditional space-invariant isotropic kernel utilized by a bilateral filter frequently leads to blurry edges and "gradient reversal" artifacts in image enhancement. The thesis presents a space-variant bilateral filter, which can preserve the edge structures with different scales and directions in image smoothing. Two error-bounded approximation methods are also used to obtain an accurate linear time acceleration method of space-variant bilateral filter. The advantages of the proposed filter is validated in applications including: image denoising, image enhancement, and image focus editing. Experimental results demonstrate that our fast and error-bounded space-variant bilateral filter is superior to state-of-the-art methods.
(3) A fast guided mesh filtering method based on minimum spanning tree and multiple output linear ridge regression is proposed. To reduce the computational cost, the existing two-stage mesh filtering methods employ a small and fixed local neighborhood in the normals filtering, and blur the edges or shape features in mesh smoothing. A fast guided mesh filter is proposed, which constructs minimum spanning tree based on face centroids and normals to obtain feature-aware implicit neighborhood defined by tree similarity weight, and utilizes multiple output linear regression to accurately estimate the uncontaminated normals. With the help of the fast minimum spanning tree aggregation method, linear time guided normal filtering is achieved. Experimental results show that the proposed method is superior to state-of-the-art methods in mesh denoising.
(4) A customized summarization method for visual data collection is proposed. In the field of computer graphics, visual data collection summarization is a commonly used tool for visualization of algorithm output. While selecting samples in visual explorations is used as a component of many existing shape-space exploration systems, it has not been systematically explored. Customized summarization of visual data collection is formulated into an integer programming problem with user constraints, and take into account elements diversity, user preference and element attribute information. A solver based on $\ell_2-$box ADMM method is proposed to efficiently obtain high-quality visual data collection summarization. An user interface is also designed to facilitate the generation of desired collection summarization. Experiments verify the superiority of the proposed visual data set summarization method in terms of speed and quality in comparing with state-of-the-art summarization methods.
|袁梦轲. 可视数据的保边平滑与总结[D]. 智能化大厦3层. 中国科学院自动化研究所,2019.|
|Files in This Item:|
|Thesis.pdf（197791KB）||学位论文||暂不开放||CC BY-NC-SA||Application Full Text|
|Recommend this item|
|Export to Endnote|
|Similar articles in Google Scholar|
|Similar articles in Baidu academic|
|Similar articles in Bing Scholar|
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.