项目名称: 鲁棒模型拟合中的关键问题研究及应用
项目编号: No.61472334
项目类型: 面上项目
立项/批准年度: 2015
项目学科: 其他
项目作者: 王菡子
作者单位: 厦门大学
项目金额: 86万元
中文摘要: 在处理实际的任务时,来自于图像或图像序列的数据可能是不精确的,几乎不可避免会被如传感器噪声、错误的特征提取、错误的特征匹配、分割误差等因素产生的离群数据所污染;而且,数据中也可能包含多个模型结构。这样,计算机视觉领域中的一个重要任务就是从含有大量噪声和多结构的数据中估计出数据中所包含的模型结构的参数,并把属于不同模型结构的数据分割开。然而至今,这依然是一个非常具有挑战性的任务。本项目针对鲁棒拟合方法中存在的不足,拟在三个方面解决参数模型拟合中存在的关键问题:有效地进行样本抽样、自适应地估计模型的内点噪声尺度、以及鲁棒地拟合和分割多结构数据。进而,我们还将把所提出的鲁棒拟合方法应用在计算机视觉任务中(包括自然场景中的文本检测等)。本研究对于提高现有模型拟合方法的鲁棒性和精确度具有重要的理论指导意义,并对计算机视觉中的实际任务(如人脸识别、三维重构、运动估计、图象分割等)有着重要的应用价值。
中文关键词: 计算机视觉;模式识别;鲁棒拟合;参数模型估计;文本检测
英文摘要: When dealing with practical tasks , data from images or image sequences may be inaccurate. It is almost unavoidable that data will be contaminated by the factors such as sensor noises, errors in feature extraction and feature matching, segmentation errors, and so on. Moreover, data may also include multiple model structures. Thus, an important task in the field of computer vision is to estimate the parameters of model structures from data containing a lot of noises and multiple structures, and segment data belonging to different model structures. So far, however , this is still a very challenging task. Considering the shortcomings of the existing robust fitting methods, this project intends to solve three key problems in parametric model fitting: effectively sample data, adaptively estimate the inlier noise scale of a model, and robustly fit and segment multi- structured data. Furthermore, we will apply the proposed robust fitting methods to computer vision tasks (including text detection in natural scenes and so on). The research in this project is theoretically important for guiding to improve the robustness and accuracy of existing robust fitting methods, and has important application values on practical tasks in computer vision (such as face recognition, 3D reconstruction, motion estimation, image segmentation, and so on).
英文关键词: Computer Vision;Pattern Recognition;Robust Fitting;Parametric Model Estimation;Text Detection