【作者】 周吉；

【导师】 王天江；

【作者基本信息】 华中科技大学 ， 计算机应用技术， 2009， 硕士

【摘要】 由于自然图象的背景复杂,在进行图象的轮廓检测时可能会有一些重要轮廓丢失,从而会给接下来对象检测和分析、对象跟踪等工作带来额外的困难。自然图象中的轮廓检测算法主要是围绕着梯度特征和曲线连续性特征进行图象的概率轮廓的检测,通过建立逻辑回归和曲线连续性模型检测和补全图象轮廓。这种轮廓检测结合补全的方法并不局限于采用某一种边缘检测算法,因此具有广泛的适用性。以亮度梯度、颜色梯度和纹理梯度为基础,采用了用逻辑回归的方法检测对象的轮廓边缘。对图象的局部区域的象素点进行亮度、颜色和纹理的梯度对比,区域差分直方图的值作为中心点的梯度值,逻辑回归模型组合这三个梯度特征,学习到特征参数,从而判断图象中每个点是轮廓的概率。考虑到自然图象中背景的复杂性,重要轮廓的丢失可能不可避免,利用曲线连续性模型对概率轮廓建模,通过学习轮廓的局部特征从而将丢失的轮廓补全。轮廓补全工作是从德劳奈三角剖分图开始的,它是沃罗诺伊图的偶图。约束的德劳奈三角剖分将图象的概率轮廓补全之后,通过曲线连续性模型学习到模型参数,从而判断约束的德劳奈三角剖分图中补全的轮廓哪些是真实的轮廓,哪些不是的,最后得到的轮廓是较完整的图象轮廓。曲线的连续性模型分为两种,一种是局部的,一种是全局的。曲线的局部连续性轮廓补全采用的是逻辑回归模型,曲线的全局连续性轮廓补全采用的是条件随机场模型。在实验中选择坎尼边缘检测算法与概率轮廓检测和补全算法做比较,分别在查全率和准确率以及全度量上进行了性能对比分析,结论表明在自然图象中采用概率轮廓检测和补全算法的性能要比坎尼边缘检测算法略优。更多还原

【Abstract】 Because of the complexity of the background in the natural images, some important boundary may be missed on detecting, which will introduce additional difficulty to the following object detecting, analysis or tracing. This work of boundary detecting in the natural images is surrounded with the gradient features and curvilinear continuity features to detect the probability of boundary in images. Boundary of images is detected and completed by constructing logistic regression and curvilinear continuity models. This method of boundary detecting combined with completion isn’t limited with any kind of edge detecting arithmetic, then it is global applicable.Based on the brightness gradient, color gradient and texture gradient, boundary of object in the image is detected using logistic regression. It is necessary to compare the gradient of brightness, color and texture of pixels with each other in the two half areas of images, and the gradient value of the center pixel is computed from the difference of histograms of the two half. After combining the three gradient features with logistic regression model and learning parameters of these features, the probability of boundary(Pb) to every pixel in the images could be decided.The missing of significant boundary might be impossible to avoid with regard to the complex of the background in natural images. Using the curvilinear continuity models with the probability of boundary (Pb), gaps in the Pb will be completed by learning the features of local boundary. Before the task of completion of boundary, it is important to building a Delaunay Triangulation(DT) with Pb, which is the antithetic graphics of Voronoi. Constrained Delaunay Triangulation(CDT) completes the Pb, and then parameters in the models are learned from the curvilinear continuity, finally, whether a completed edge in the CDT is a real bound or not could be fixed. As a result, the output boundary is better in its integrity. There are two curvilinear continuity models. One is local, while the other is global. The local curvilinear continuity boundary completion uses logistic regression model, while the global curvilinear continuity boundary completion uses Conditioned Random Field (CRF) model.In the experiment, the arithmetic of detecting probability of boundary and its completion arithmetic are compared with Canny detector. The comparison aspects include ratio、precision and F-measure. The result suggests that using the arithmetic of detecting and completing probability of boundary is better than Canny edge detecting arithmetic in performance.更多还原