【摘要】：如何在場景復雜的圖片中快速得到用戶感興趣的目標,正成為如今計算機視覺和模式識別領域的熱點和難點。圖像分割作為一種提取目標的有效途徑一直以來都受到學者們的廣泛關注,也已經取得了較多的研究成果。但是依舊存在較多的難點,比如:圖像包含的信息越來越多,圖像的特征也越來越豐富,單一的特征已經不能滿足如今的技術要求;在分割粒度方面,基于像素級的分割框架往往會導致分割目標的不完整性,而一些基于區(qū)域的分割框架則存在細節(jié)方面的丟失,并且較為依賴預分割區(qū)域的準確性。本文針對圖像的多源屬性協(xié)同分割問題,主要進行了如下幾個方面的創(chuàng)新工作:首先,提出了一種有效的紋理建模方式。通過對傳統(tǒng)的多尺度結構張量進行精簡和非線性濾波處理得到非線性精簡多尺度結構張量,并將非線性精簡多尺度結構張量與全局變分流結合構成我們所使用的紋理描述子。采用圖割框架對該紋理描述子的有效性進行實驗,通過與常用紋理特征的實驗對比驗證了該紋理描述子在紋理描述力上有較好的效果,而且其較低的維度特質也給后續(xù)的概率建模帶來了效率上的提升。其次,提出了基于虛擬節(jié)點的圖像多源屬性的協(xié)同分割框架,并將該框架用于多特征的協(xié)同分割。以L*a*b顏色特征和我們所提出的紋理建模方式為例對基于虛擬節(jié)點的顏色紋理協(xié)同分割方法進行了實驗。通過與傳統(tǒng)的顏色紋理能量混合模型和單一的特征的分割結果進行對比,說明了該分割框架能夠較好的吸納不同特征分割中的優(yōu)勢部分。最后,提出了基于上下文信息的圖像多源屬性的協(xié)同分割框架,并將該框架用于多粒度的協(xié)同分割。以邊緣增強的均值漂移算法得到的同質區(qū)域粗粒度和原像素細粒度為例對基于上下文信息的粗細粒度協(xié)同分割方法進行了實驗。通過與基于單一粒度的分割結果進行對比,說明該分割框架能夠在細節(jié)和目標整體性上面表現(xiàn)良好。本文通過大量的仿真實驗驗證了本文紋理建模方式、基于虛擬節(jié)點的多特征協(xié)同分割和基于上下文信息的多粒度協(xié)同分割的實效性和可用性,并具有良好的應用前景。
[Abstract]:How to quickly get users interested in the complex scene is becoming a hot and difficult point in the field of computer vision and pattern recognition. As an effective way to extract the target, image segmentation has been widely concerned by scholars and has also taken more research results. There are many difficulties, such as: more and more information is included in the image, and the features of images are becoming more and more rich. The single feature can not meet the technical requirements of today. In the aspect of segmentation granularity, the segmentation framework based on pixel level often leads to the incompleteness of the segmentation target, and some segmentation frameworks based on the region exist in detail. For the problem of multi source attribute synergetic segmentation, this paper focuses on the following aspects: first, an effective texture modeling method is proposed. The nonlinear precision is obtained by the traditional multi-scale structure Zhang Liangjin row simplification and nonlinear filtering. The texture descriptor of the texture descriptor used by the nonlinear simplification of multi scale structure tensor and global variation is introduced. The validity of the texture descriptor is tested by the graph cut frame. The texture descriptor is better than the common texture descriptor. The effect, and its lower dimensional characteristics also brings efficiency to the subsequent probabilistic modeling. Secondly, a collaborative segmentation framework based on the multi source attributes of virtual nodes is proposed, and the framework is used for multi feature synergetic segmentation. The L*a*b color feature and the texture modeling method we put forward are based on virtual nodes. By comparing with the traditional color texture energy mixing model and the single feature segmentation result, it shows that the segmentation framework can well absorb the advantages of different feature segmentation. Finally, the cooperative segmentation of multi source attributes based on context information is proposed. The frame is cut and the framework is used in multi granularity cooperative segmentation. The coarse granularity and fine grain size of the homogeneous region obtained by the edge enhanced mean shift algorithm is used as an example to experiment on the coarse and fine granularity cooperative segmentation method based on context information. By comparing the segmentation results based on the single granularity, the segmentation framework is illustrated. Through a large number of simulation experiments, this paper validates the texture modeling method in this paper, the effectiveness and availability of multi granularity cooperative segmentation based on virtual nodes and multi granularity based on context information, and has a good application prospect.
【學位授予單位】：華中科技大學
【學位級別】：碩士
【學位授予年份】：2016
【分類號】：TP391.41

【參考文獻】

相關期刊論文 前2條

1 韓守東;趙勇;陶文兵;桑農;;基于高斯超像素的快速Graph Cuts圖像分割方法[J];自動化學報;2011年01期

2 劉麗;匡綱要;;圖像紋理特征提取方法綜述[J];中國圖象圖形學報;2009年04期

，

本文編號：2157142