基于图像边缘对比的显著性检测算法

发布时间：2018-04-06 05:39

本文选题：显著性检测　切入点：超像素　出处：《吉林大学》2017年硕士论文

【摘要】：目标(即对象、人、像素等)的显著性是相对于其邻居突出的状态或特性。显着性检测被认为是一个关键的注意力机制,用以促进学习和生存。通过使有机体将其有限的感知和认知资源集中在可用的感应数据的最相关的子集,有机体可以利用最少的能量获取最多的知识。显著性通常源自项目及其邻域之间的对比,诸如由白点围绕的红点,应答机的闪烁消息指示符,或者在其他安静环境中的大噪声。显性检测通常在视觉系统的上下文中进行研究,但类似的机制在其他感应系统中操作。什么是显著的这个定义可以通过接受训练的方式来解决:例如,对于人类受试者,特定的信息可以通过训练变得突出。当注意力的部署由显著性刺激驱动时,它被认为是自下而上,无记忆和反应性的。注意机制也可以由自上而下,记忆相关或预期的机制来指导,例如在越过街道之前向前看运动对象或侧向。人类和其他动物难以同时注意多个项目,因此他们面临着不断地整合和优先考虑不同的自下而上和自上而下的影响的挑战。显著性检测算法已经成为CV领域一项实用的工具。目前所有的显著性检测算法可以分成两个种类:自顶向下的检测算法和自底向上的检测算法,目前的主流依旧是自底向上的检测算法。在这篇论文中,本文提出一个新颖的自底向上的方法:基于边缘对比度(boundary prior,边缘先验知识)的图像显著性检测算法。本文提出的方法在便于实现的同时,也保证了检测的质量。本文通过以下几步来获得超像素的分割:首先,利用超像素算法——SLIC算法,对被检测的源图像进行预分割,因而可以将它分割成一系列紧密均匀一致的超像素块。第二步,提取相应的超像素块的一系列特征值。第三步,为了强调距离当前超像素块更近的边缘超像素块的作用,本文继而计算每个超像素两个特征值:一个是边缘不相似性的值,另一个是超像素块到边缘超像素块的最短路径。第四步,通过计算,将超像素不相似性和到边缘超像素的最短路径长度值融合在一起,从而计算出每个超像素属于背景区域的概率。第五步,通过一系列的最优化方法优化显著性概率值,得到最终的显著值结果,本文也将展示如何推导这个最优化公式。最后,将超像素的值赋予相应的像素点,得到最终的显著图。最后通过大量的对比实验来证明算法的有效性和优秀效果。
[Abstract]:The salience of a target (i.e. an object, a person, a pixel, etc.) is the state or characteristic that stands out relative to its neighbor.Salience testing is considered to be a key attention mechanism to promote learning and survival.By enabling organisms to concentrate their limited perceptual and cognitive resources on the most relevant subsets of available sensing data, organisms can acquire the most knowledge with the least amount of energy.Salience is usually derived from comparisons between items and their neighbors, such as red dots surrounded by white dots, flickering message indicators of transponders, or large noises in other quiet environments.Explicit detection is usually studied in the context of visual systems, but similar mechanisms operate in other sensing systems.This definition of what is significant can be solved by training: for example, for human subjects, specific information can become prominent through training.When attention deployment is driven by significant stimuli, it is considered bottom-up, memoryless and reactive.Attention mechanisms can also be guided by top-down, memory-related or anticipated mechanisms, such as looking forward or sideways before crossing the street.Humans and other animals have difficulty paying attention to multiple projects at the same time, so they face the challenge of continuously integrating and prioritizing different bottom-up and top-down impacts.Significance detection algorithm has become a practical tool in CV field.At present, all significant detection algorithms can be divided into two categories: top-down detection algorithm and bottom-up detection algorithm, the current mainstream is the bottom-up detection algorithm.In this paper, a novel bottom-up approach is proposed: an image salience detection algorithm based on edge contrast boundary prior.The method proposed in this paper is easy to realize and also ensures the quality of detection.In this paper, the super-pixel segmentation is obtained by the following steps: firstly, the source image is presegmented by using the super-pixel algorithm (SLIC), so that it can be divided into a series of compact uniform superpixel blocks.In the second step, a series of eigenvalues of the corresponding hyperpixel blocks are extracted.Third, in order to emphasize the role of edge superpixel blocks closer to the current superpixel blocks, this paper then calculates two eigenvalues for each superpixel: one is the edge dissimilarity.The other is the shortest path from the hyperpixel block to the edge superpixel block.In the fourth step the super-pixel dissimilarity and the shortest path length to the edge super-pixel are fused together so as to calculate the probability of each super-pixel belonging to the background region.In the fifth step, through a series of optimization methods to optimize the significant probability value, the final significant value result is obtained. This paper will also show how to deduce this optimization formula.Finally, the value of the super pixel is assigned to the corresponding pixel, and the final salience diagram is obtained.Finally, through a large number of comparative experiments to prove the effectiveness and excellent effect of the algorithm.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】