基于多尺度区域对比的视觉显著性检测算法研究

	基于多尺度区域对比的视觉显著性检测算法研究
其他题名	Visual Saliency Detection Algorithm Based on Multi-scale RegionContrast
	成培瑞
学位类型	硕士
导师	王建立
	2015-10
学位授予单位	中国科学院大学
学位专业	机械电子工程
关键词	视觉显著性显著性计算模型图像分割多尺度区域对比
摘要	随着互联网和各种高科技电子设备的出现和发展，图像和视频作为一种新兴的信息载体，其数量呈现爆炸式地增长。但现在的计算机对图像和视频的处理速度并不能跟上其数量的增长速度，因此为了提高计算机对视觉数据的处理能力，基于人类视觉系统机制的视觉显著性这一概念应运而生。视觉显著性是基于对人类视觉系统机理的研究而被提出的，可以被认为是图像中的元素吸引人类视觉注意力的能力。视觉显著性检测算法正是对这种能力的量化，通过检测算法得到图像中的显著性分布，并将显著区域与背景分离，以便对显著区域单独分析处理，从而提高计算机对视觉数据的处理速度。本文的主要研究内容是基于多尺度区域对比的视觉显著性检测算法。算法基于自底向上的视觉选择性注意机制，利用基于多尺度的图像分割、基于区域的颜色特征对比和视觉规律检测到边缘清晰、内容完整的图像显著区域。本文围绕上述研究内容，提出了基于多尺度的超像素图像分割和基于多尺度区域对比的显著性检测算法，主要内容如下： (1)参照现有的基于超像素图像分割算法原理，提出基于多尺度的超像素图像分割。首先采用SLIC 算法分割图像生成超像素，然后以超像素为单位聚类以生成面积更大的超像素，其中超像素间特征距离由其颜色直方图距离表示，并基于超像素大小确定超像素特征对比范围，降低计算复杂度。最后迭代计算新中心点并重复聚类直到特征残差满足给定要求。通过此方法可生成代表不同尺度的不同大小的超像素，用作后续的基于多尺度区域对比的显著性检测算法中的区域表示。 (2)为了对图像中的显著目标进行更精确的检测，提出一种基于多尺度区域对比的视觉显著性检测算法。首先利用基于多尺度的超像素图像分割将图像分别分割为不同数目的超像素，对超像素内的像素颜色值取平均从而提取超像素的颜色特征；然后在单一尺度下根据显著特征的稀少性计算超像素的稀少性表示，根据显著特征的空间分散性及移动视觉焦点计算超像素的空间分散性表示，并将两种表示融合得到该尺度下的子显著图；最后通过取各尺度超像素显著值的平均值融合多尺度生成最终显著图。实验证明，以MSRA 图像数据库中1000 张随机图片为例，该算法较现有效果较好的RC(Region-Contrast)算法，显著目标识别的精确率提高了14.8%，F-Measure 值提高了9.2%。与现有算法相比，本文算法提高了算法对显著目标大小及背景复杂程度的适应性，减少了背景对显著目标识别的干扰，具有更好的一致性。
其他摘要	With the appearance and development of Internet and various kinds of electronic appliances, the number of images and videos increases explosively as a novel information carrier. However, the speed of processing images and videos cannot catch up with its increasing speed. Therefore, in order to improve the ability of processing visual data for computer, the concept of visual saliency come up based on human vision system. Vision saliency is brought based on the human’s visual sensing mechanism, and can be described as the ability to attract attention for elements in the scene. The ability is quantified by the algorithm for detection of visual saliency. We can get the distribution of saliency in the image through the algorithms, and separate salient region from background for separate analysis and processing to salient region, therefore increase the processing speed for visual data. The main research of this paper mainly focuses on the visual saliency detection lgorithm based on multi-scale region comparison. Based on from-bottom-to-top visual selective focus mechanism, the algorithm can detect sharp-edge and intact-content salient region with the method of multi-scale image segmentation, comparison of region color’ characters and human vision’s regular pattern. Based on above discussion, we propose the multi-scale super-pixel image segmentation algorithm and visual saliency detection algorithm based on multi-scale region comparison. The main content is listed below: (1) on the basis of the existing super-pixel segmentation algorithm, we propose the super-pixel image segmentation algorithm. At first, generate super-pixels with SLIC algorithm segmenting the image, and then cluster with the unit of super-pixel to create super-pixels in higher scale, in which the super-pixels’ character distance is presented by the distance of their color histograms, and lower the computation complexity by determining the super-pixels’ characters comparison region based on the size of super-pixel. Iteratively compute new centers and repeat clustering until the residual error of characters satisfy the threshold. In this way, we can get multi-size and multi-scale super-pixels for the region representation in the following algorithm for visual saliency detection based on multi-scale region comparison. (2) In order to carry on more precise detection for salient target in the image, we propose an algorithm for visual saliency detection based on multi-scale region comparison. Firstly, the multi-scale super-pixel image segmentation algorithm separates the image into various numbers of super-pixel, and averages the color values in the super-pixel to subtract the color character of the super-pixel. Secondly, compute the rarity of super-pixel in single scale based on the salient character, and also compute the dispersibility of the super-pixel based on dispersibility of salient character in space and moving visual focus, and merge the rarity and dispersibility to get the sub salient diagram in this scale. Finally, generate final salient diagram by averaging the salient values of super-pixels on every scale. The result shows that taking the 1000 random images int MSRA image database, the proposed algorithm improves the accuracy of salient target recognition by 14.8%, F-Measure by 9.2% compared with existing well-performed RC(Region-Contract) algorithm. Compared with existing algorithms, the proposed algorithm improves the adaptability of the algorithm for the size of salient target and complexity of the background, reduces the disturbance of the background for salient target recognition, and has better consistency.
语种	中文
文献类型	学位论文
条目标识符	http://ir.ciomp.ac.cn/handle/181722/49255
专题	中科院长春光机所知识产出
推荐引用方式 GB/T 7714	成培瑞. 基于多尺度区域对比的视觉显著性检测算法研究[D]. 中国科学院大学,2015.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
成培瑞.pdf（2144KB）	学位论文		开放获取	CC BY-NC-SA	请求全文