基于形状先验的变分稀疏分割模型研究

发布时间：2018-03-21 23:01

本文选题：目标分割　切入点：形状先验　出处：《浙江大学》2017年博士论文　论文类型：学位论文

【摘要】：利用计算机模拟人眼的功能准确分割图像中的目标,对于计算机视觉和图像处理至关重要。由于在实际图像中,目标可能受到噪声、遮挡、缺损或背景粘连等因素的影响,仅依赖图像本身的信息对目标进行分割往往效果不佳。基于形状先验生成过完备形状字典,并结合稀疏表示模型对目标分割过程进行监督可较好的改善分割效果。本文在深入分析了国际上一些经典的稀疏形状表示模型和变分分割方法的基础上,重点研究了形状的稀疏表示方法、映射空间过完备字典的生成方法以及结合稀疏表示的变分目标分割方法等内容,提出了多种基于形状先验的变分稀疏分割模型。论文的主要工作和创新之处在于:1.基于相关命题推出通过扩展形状映射系数的定义,可基于映射公式构造形状凸集和其稀疏子集。利用上述子集中的稀疏组合约束输入目标在形状凸集上的投影,可进一步建立基于映射空间的变分稀疏分割模型。论文对结合独立形状分量和非线性核空间的凸形状集研究表明,上述基于映射凸集的变分稀疏形状表示方法具有一定的通用性,该方法也为建立映射空间变分稀疏分割模型提供了一种新的思路。2.结合变分水平集方法、独立形状分量和稀疏形状表示,提出了一种基于稀疏独立分量表示的变分目标分割方法。该方法解决了直接使用独立分量构造稀疏形状组合无法确保重建形状的有效性问题。此外,模型还使用对数极坐标形状代替原始形状,实现了稀疏形状表示模型中的目标自动对齐。通过在独立形状分量张成的空间中构造映射形状集,并迭代求解水平集函数在该形状集上的稀疏表示等方法,解决了利用稀疏独立分量表示模型监督变分目标分割的问题。3.本文结合近期发展起来的“Wake-sleep”优化方法,提出了一种分层分割框架。该框架将能量最小化过程分解为“Wake”和“Sleep”两个阶段,通过“Wake”阶段强化模型对底层图像的表达。同时,通过“Sleep”阶段强化高层能量对输入形状的表示。此外,本文还提出一个对偶连接能量项来平衡“Wake”和“Sleep”阶段的演化。综合上述几项技术,本文构建了一种基于隐含核稀疏形状表示的变分分割模型,该模型解决了非线性核空间中稀疏近邻的搜索问题及利用隐含核形状表示指导底层变分目标分割的问题。4.针对小样本情况下过完备字典形状表示能力弱的缺陷,提出了一种基于映射字典群的变分稀疏分割模型。该模型在传统稀疏表示函数中引入了局部约束项和局部权重系数来实现形状的局部调整。与传统方法相比,新的模型充分挖掘了训练集中的形状信息,大幅提高了基于形状先验的变分稀疏分割模型在小样本情况下的形状表达能力,解决了目标存在局部形变的情况下,现有变分稀疏分割方法无法充分利用训练集样本的局部特征重建目标形状的问题。5.提出了距离约束概率形状和对数极坐标模糊分解两种形状建模方法,改善了传统稀疏组合形变能力不足的问题。其中,基于距离约束的概率形状通过引入距离约束的概念模拟了平均形状普遍存在的“过渡带”特征,拓展了单个形状样本的内涵。基于对数极坐标模糊分解的形状字典生成方法对训练集形状进行了更深层次的分解和挖掘,为小样本情况下构建过完备形状字典群提供了更加丰富的信息,大幅改善了小样本情况下的分割精度。论文在公开数据集和自建数据集上的实验结果表明,上述基于形状先验的变分稀疏分割模型具备较好的分割精度和噪声鲁棒性。
[Abstract]:Accurate segmentation of the object in the image using computer to simulate human eyes features for computer vision and image processing is very important. Due to the actual image, the target may be affected by noise, occlusion, effect of defect or background adhesion and other factors, the target segmentation is often ineffective only depend on the image itself. Based on the prior shape generates a complete shape the dictionary, and combined with the sparse representation model of target segmentation process supervision can better improve the segmentation effects. Based on the in-depth analysis of the shape of some classical sparse representation based on the model and variational segmentation methods, focusing on the shape of the sparse representation method, method of generating space mapping over complete dictionary and the sparse representation the variational target segmentation method, put forward a variety of shape prior variational sparse image segmentation model based on the main. The work and innovations: 1. based on the related proposition launched by the expansion of the definition of shape mapping coefficient, mapping formula is constructed and its convex shape based on sparse subset. Sparse constrained combinatorial input target using the subset in the shape of projection on convex set, can further build mapping space variational image segmentation model based on the sparse. The shape of convex shape with independent component and nonlinear kernel space research show that the mapping of convex variational method based on sparse representation of shape has certain versatility, this method also sparse segmentation model provides a new method combining.2. with variational level set method for establishing the mapping space, independent component shape and sparse shape representation, the paper proposes a segmentation method based on sparse representation independent component variational target. The method solves the problem of directly using independent component structure sparsed To ensure the effectiveness of the combination of shape reconstruction of shape. In addition, the model also uses log polar shape instead of the original shape, the shape of said automatic target alignment sparse model. By constructing a map shape in a shape independent component space set, and iterative solution of the level set function in the shape of the sparse set. Said method, it solves the sparse representation model of independent component supervision variational target segmentation problem of.3. combined with the recent development of "Wake-sleep" optimization method, proposes a hierarchical segmentation framework. The framework of the energy minimization process is decomposed into "Wake" and "Sleep" two stages, enhanced the expression of the underlying model the image through the "Wake" stage. At the same time, through the "Sleep" stage to enhance the senior energy input shapes. In addition, this paper also offers a dual energy connection To balance the evolution of the "Wake" and "Sleep" stage. The above technique, this paper constructs a sparse hidden nuclear shape representation based on variational segmentation model, the model solves the problem of nonlinear kernel space search and sparse neighbor by hidden nuclear shape representation to guide the underlying variational segmentation problem.4. in the case of small samples over complete dictionary weak shape representation, we propose a segmentation model based on sparse dictionary mapping group. The model indicates that the partial adjustment function was introduced in the local constraints and local weights to achieve the shape in the traditional sparse. Compared with traditional methods, the new model fully tap the shape information in the training set, a substantial increase in the ability to express the shape of variational sparse segmentation model in the case of small samples based on the shape of the target, to solve the existence of local shape Under the condition of variable,.5. local feature reconstruction target shape problem existing variational sparse segmentation method can not make full use of the training sample set the distance constraint probability shape and log polar fuzzy decomposition two shape modeling method, improved the traditional sparse combination deformation capacity problem which, based on the probability of shape by distance constraint the introduction of the concept of distance constraint to simulate the average shape of widespread "transition zone" feature, expand the connotation of individual shapes. The shape of sample log polar fuzzy dictionary generation method based on the decomposition of the training set shape is decomposed and deeper, for the case of small samples to construct overcomplete dictionary group provides shape more information is greatly improved under the condition of small samples. The accuracy of segmentation in the public data sets and built experimental results on the dataset table It is clear that the proposed variational sparse segmentation model based on shape prior has better segmentation precision and noise robustness.

【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】