图像区域显著性方法研究

发布时间：2018-03-07 21:04

本文选题：面向应用的显著性　切入点：区域显著性　出处：《西北工业大学》2016年博士论文　论文类型：学位论文

【摘要】：图像特征区域检测是计算机视觉与模式识别中的重要研究方向,具有多年的研究历史。作为多研究领域的基础步骤,特征区域检测的质量直接影响了后续应用的算法结果。由于传统的特征区域检测方法与后续应用相对独立,因此存在着难以根据应用场景及数据集变化而自适应选择特征等问题。受人类视觉显著性模型的启发,本文将显著性引入到特征区域检测研究中来,并针对计算机视觉中多个广泛应用的领域,例如特征匹配、目标识别等,提出基于特定应用场景的区域显著性算法,该算法用于估计图像中各区域的显著性值。多类型图像上的实验结果证明了,显著性高的区域比显著性低的区域对完成后续应用有着较好的算法效果。作为基础研究领域,区域显著性研究对提高特征区域在其他领域的应用效果等方面意义重大。本文的主要内容和创新点包括:针对传统特征区域检测算法与区域匹配率相脱节的问题,本文提出了一种基于特征匹配的区域显著性方法,该方法将区域的匹配率引入到区域的显著性衡量中来,使得高显著性区域比低显著性区域具有较高的正确匹配率;该方法还较好的解决了自然场景图像中多维特征概率分布建模的问题。所提方法定义区域的显著性为该区域正确匹配的概率分布与错误匹配概率分布之间的Kullback-Leibler(K-L)散度,并基于自然场景图像中区域特征的统计特性,利用椭圆对称分布和高斯对数分布对多维特征概率密度函数进行了建模,给出了基于K-L散度的区域显著性计算方法。在多个自然场景数据集上的实验结果验证了所提算法的有效性和优越性,同时实验结果表明特定区域的显著性可随训练数据集进行自适应的调整。针对现有区域显著性计算方法存在的多维特征分布建模时概率模型受限的问题,本文提出了一种基于离散样本K-L散度估计的区域显著性方法,该方法可适用于具有不同统计特征的图像数据集,比现有方法适用范围较广,计算速度较快。所提方法从以下两个方面对现有方法进行了改进:1)使用随机样本直接估计基于K-L散度的区域显著性,避免对区域特征的概率分布添加过强的限制条件;2)优化现有方法计算步骤,省略了显著性计算中的中间步骤。本文将该方法应用至自然图像、纹理图像和仿真数据集等多类图像数据集以及不同特征描述方法,实验结果证明,该方法可有效地计算不同类图像区域的显著性;具有较广泛的适用性。针对基于匹配的目标识别,本文提出了一种综合考虑图像区域的表示性、区分性以及匹配有效性的显著性方法,该方法较现有方法可获得较高的目标识别正确率。所提方法中区域的表示性指该区域为感兴趣目标的一个特征区域,可描述目标整体或局部;将区分性区域定义为存在于感兴趣目标中且较少在背景中出现的区域,利用基于离散样本的K-L散度估计方法计算区域的区分性值;并将匹配率引入到区域的显著性衡量中,定义具有较高区分性和正确匹配率的目标区域为高显著性区域。在多目标识别数据集上的实验结果表明,所提方法提取的高显著性区域可区分复杂背景中的感兴趣目标,具有较高的正确识别率,并有助于获得较准确的定位和姿态估计结果。
[Abstract]:Image feature detection is an important research direction of computer vision and pattern recognition in the study of history has many years. As a basic step in the research field of quality characteristics, region detection directly affects the subsequent application of the algorithm results. Because the traditional method of detecting feature area and subsequent application of relatively independent, so there are not according to the application scenarios and data set change and adaptive feature selection problems. Inspired by the human visual saliency model, this paper was introduced to study detection feature regions, and for the more widely used in computer vision field, such as feature matching, object recognition, put forward specific application scenario of the region saliency based algorithm the algorithm is used to estimate the significance of each area in the image. The value of multi type images on the experiment results show that was more significant than the region Low area has a good effect on the algorithm complete follow-up application. As the basic research, research of salient regions is of great significance to improve the effectiveness of regional characteristics in other areas and so on. The main contents and innovations of this paper include: according to the characteristics of the traditional region detection algorithm and area matching rate of the separation, is proposed in this paper. A significant regional feature matching method based on the method of regional matching rate into the saliency measure to make significant regional high correct matching ratio is higher than low saliency rate; the method also solves the problem of multidimensional probability distribution modeling problems in natural scene images the probability distribution and error significantly. The proposed method defines the area of the region the correct matching probability distribution between Kullback-Leibler (K-L) divergence, and based on the natural The statistical characteristics of regional characteristics of the scene image, using elliptical symmetry and logarithmic distribution of Gauss multidimensional probability density function for modeling, calculating K-L divergence region saliency based method. In a number of natural scene data sets. The experimental results verify the validity and superiority of the proposed algorithm, and experiment the results show that significant specific area with the training data set for adaptive adjustment. The probability distribution model of multidimensional constrained model for the existing region saliency calculation method of the problem, this paper proposes a method of discrete sample region saliency estimation based on K-L divergence, this method can be applied to image data with different statistics the feature sets, a wider scope than the existing method, fast calculation speed. The proposed method from the following two aspects to improve the existing method: 1). With a random sample of direct estimation of K-L divergence region saliency based on probability distribution characteristics of the region to avoid adding too strong restrictions; 2) to optimize the existing method of calculation steps, omitted significant in the calculation of the intermediate steps. This method was applied to natural images, texture images and simulated data sets, and other types of the image data sets and different feature description method, experimental results show that this method can effectively calculate the significant different image regions; has a wide applicability. For target recognition based on matching, this paper proposed the image area of a comprehensive consideration, distinguishing and matching significantly the effectiveness of the method this method, compared with the existing method can achieve high recognition accuracy. The proposed method in regional representation refers to a feature region in the region as the target of interest, which can describe the overall goal or The local area is defined as; distinction exists in the target of interest and less in the background area, using K-L divergence estimation of discrete sample discrimination calculation method based on the value of the region and into the matching rate; saliency measure, defined with target area area and high correct matching rate as a high saliency. Multi target recognition data sets. The experimental results show that the proposed method to extract high saliency regions can be distinguished in the complex background of interested target, has higher correct classification rate, and helps to achieve accurate positioning and attitude estimation results.

【学位授予单位】：西北工业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】