显著性目标分割模型及其应用

发布时间：2018-07-06 07:21

本文选题：显著性目标分割 + 图像分割　；参考：《深圳大学》2016年博士论文

【摘要】：视觉感知在人类获取周围环境的知识中起到了不可替代的作用,而图像是视觉感知的重要载体之一。随着智能手机的广泛普及,互联网上的图像资源呈现爆炸性的增长。如果可以模拟人类的视觉注意感知机制,对进入视野的海量图片信息自动找出最重要、最显著的部分进行处理,这样有助于突破信息处理的瓶颈,节省计算和分析资源,提高信息处理的速度。显著性目标就是一个场景中最重要、最吸引人注意的部分。显著性目标分割就是自动将输入图片中最显著的目标进行精确分割的一种模型。近年来,显著性目标分割模型吸引了研究者极大的兴趣,涌现出了众多的优秀的模型与算法。而且显著性目标分割技术在机器人视觉、视频安全监控、目标探测与识别、视频压缩与编码、运动检测、视频目标跟踪等领域得到了广泛应用。本文从基于区域的图像分割研究出发,逐步深入到显著性目标分割问题,结合子空间聚类理论、非局部各向异性扩散方程、流扩散等相关理论,提出了一些新的算法和模型。本文工作主要包括以下几个方面：1.在基于区域的图像分割问题研究中,提出了一个新的基于非凸半范数约束的稀疏低秩耦合的图像分割模型。将子空间聚类模型引入到图像分割问题中,并且采用S1/2范数替代核范数来近似低秩表示,采用L1/2范数替代L1范数来近似稀疏表示。研究发现,对低秩表示问题和稀疏表示问题而言,S1/2范数和L1/2范数虽然是非凸的,但它们是一种更好的松弛,且它们的极小化问题具有闭式解。通过ALM算法分离变量,利用半阈值算子得到一种有效的离散算法。大量的实验证明了模型分割的准确性和对噪声的鲁棒性。2.进一步深入研究了非凸低秩子空间聚类问题。用Schatten-q拟范数松弛低秩子空间聚类模型中的秩函数,提出了基于Schatten-q正则化的低秩子空间聚类模型。我们证明了Schatten-q正则最小化问题的最优解可以通过一个广义的矩阵软阈值收缩算子表示,模型通过自适应线性交替方向乘子法(LADMAP)求解。理论分析证明,广泛应用的软阈值(SVT)算法和半阈值算法分别是我们新模型在q=1和q=0.5的两个特例。大量的实验证明了我们方法的有效性。3.本文进一步研究了显著性目标分割问题。本文将视觉显著性目标检测过程建模为一个非局部各向异性扩散方程,提出了一个基于非局部各向异性扩散方程的两阶段显著性检测模型。在第一阶段,利用边界先验进行了由边界到全图像的背景扩散,通过排序的方法选择了部分显著性值高的元素作为初始显著性种子。在第二阶段,从显著性种子出发将显著性值扩散到全图像域。在第二阶段的显著性迭代扩散过程中对显著性种子作了更新和优化,消除了错误种子的的影响,从而有效抑制了背景。对于模型的求解,本文提出了多方向离散格式来求解新模型。实验结果表明,无论从主观分析,还是从客观评价,新方法在3个常用数据库上都得到了良好的检测效果。4.通过对现有的基于扩散思想的显著性模型分析发现,现有模型只考虑了点与点之间的扩散强度,而没有考虑到扩散方向。这会导致,两个区域如果特征相似而所处的类别不同时,模型检测会出现较大的误差。基于这一观察,本文对前一个模型进行了改进,引入了非局部扩散张量,以在控制扩散强度的同时考虑扩散方向问题。在主方向上,抑制扩散从而保证了背景和前景之间的差异性；在其它方向上,新模型提升了扩散,从而保证相似部分作为一个整体高亮检测出来。大量实验证明了新模型的良好性能。5.进一步,本文利用多特征的并行框架来模拟人类视觉感知机制,提出一个新的基于多特征融合的显著性流模型。本文提出一个基本共识,即显著性目标分割时,在大部分特征通道中都显著的区域才认为是显著性目标。基于这个共识,在每个特征通道中,采用显著性流扩散的方法进行建模。然后用二次规划的方法训练以得到融合权重。以调整过的权重进行线性融合来得到最终的检测结果。经过大量的实验证明我们的模型有很好的综合性能。
[Abstract]:Visual perception plays an irreplaceable role in the knowledge of human access to the surrounding environment, and image is one of the important carriers of visual perception. With the widespread popularity of smart phones, the image resources on the Internet are explosively growing. If we can simulate human visual perception mechanism, the mass picture letter that enters the field of vision Interest automatically finds the most important and most significant part of the process, which helps to break through the bottleneck of information processing, save computing and analysis resources, and improve the speed of information processing. The significant target is the most important and most attractive part of a scene. The significant target segmentation is the most significant target in the input of the picture. A model for accurate segmentation. In recent years, the significant target segmentation model has attracted the great interest of the researchers, and many excellent models and algorithms have emerged. Moreover, the significant target segmentation technology is in robot vision, video security monitoring, target detection and recognition, video compression and coding, motion detection, video target tracking. Based on the study of region based image segmentation, this paper goes deep into the problem of significant target segmentation, and proposes some new algorithms and modes combining the subspace clustering theory, non local anisotropic diffusion equation, flow diffusion and other related theories. This paper mainly includes the following aspects: 1. In the study of region based image segmentation, a new image segmentation model based on sparse and low rank coupling based on non convex semi norm constraints is proposed. Subspace clustering model is introduced into the image segmentation problem, and S1/2 norm is used instead of kernel norm to approximate low rank representation, and L1/2 norm is used instead of L1 norm to approximate sparse representation. It is found that, for the problem of low rank representation and sparse representation, the S1/2 norm and the L1/2 norm are not convex, but they are a better relaxation, and their minimization problem has a closed solution. A valid discrete algorithm is obtained by using the semi threshold operator by the ALM algorithm. A large number of experiments prove the model points. The accuracy of cutting and the robustness to noise.2. further studies the problem of non convex and low rank subspace clustering. Using the rank function in the Schatten-q quasi norm relaxation low rank subspace clustering model, a low rank subspace clustering model based on the Schatten-q regularization is proposed. We prove the optimal solution of the Schatten-q regular minimization problem. By means of a generalized matrix soft threshold contraction operator, the model is solved by the adaptive linear alternating direction multiplier (LADMAP) method. The theoretical analysis shows that the widely used soft threshold (SVT) algorithm and the semi threshold algorithm are two special examples of our new model in q=1 and q=0.5. A large number of experiments prove the effectiveness of our method.3. This paper further studies the problem of significant target segmentation. In this paper, the visual saliency target detection process is modeled as a nonlocal anisotropic diffusion equation, and a two stage salience detection model based on nonlocal anisotropic diffusion equation is proposed. In the first stage, the boundary prior to the full image is carried out with the boundary prior. In the second stage, the value of saliency is spread to the whole image domain. In the second stage of the significant iterative diffusion process, the significant seeds are updated and optimized to eliminate the effect of the wrong seeds. For the solution of the model, a multi direction discrete scheme is proposed to solve the new model. The experimental results show that the new method, both from the subjective analysis and the objective evaluation, has got a good detection effect on 3 common databases through the analysis of the existing significant model based on the existing diffusion thought. It is found that the existing model only takes into account the diffusion strength between point and point, but does not take into account the diffusion direction. This will lead to a larger error in the model detection in the two regions if the characteristics are similar and the categories are different. Based on this observation, the previous model is improved and the non local diffusion tensor is introduced to the model. The diffusion strength is controlled and the diffusion direction is considered at the same time. In the main direction, the diffusion is suppressed to ensure the difference between the background and the foreground; in other directions, the new model improves the diffusion, thus ensuring the similar part as a whole bright detection. A large number of experiments prove that the good performance of the new model is further.5., Wen Li uses multi feature parallel framework to simulate human visual perception mechanism, and proposes a new model of saliency flow based on multi feature fusion. This paper proposes a basic consensus that significant regions in most characteristic channels are considered explicit targets when significant targets are segmented. Based on this consensus, each feature is connected. In the channel, the model is modeled using the method of significant flow diffusion. Then the fusion weight is trained by two programming methods. The final results are obtained by the linear fusion of the adjusted weights. A lot of experiments show that our model has good comprehensive performance.
【学位授予单位】：深圳大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】