图像内容显著性检测的理论和方法研究

发布时间：2018-11-12 10:31

【摘要】：人类的视觉系统可以在广大的、复杂的动态及静态场景中快速定位出最吸引注意的内容,这种能力被称为显著性检测。吸引注意的内容因此被叫做显著性(内容),一般来说显著性的某种特征与周围有很大差异,例如某些危险警示标志。视觉系统的显著性检测能力可以使我们更聚焦在视觉场景中的某一局部,而对场景中的其它背景部分视而不见,从而当我们面对外界的刺激时可以优先处理部分信号并更快速的做出反应。随着图像视频获取捕捉设备的发展,数据规模越来越大、内容越来越复杂,因此早期的计算机视觉算法逐渐不能胜任现如今的任务。所以,人们试图设计算法模拟人类视觉系统的显著性检测能力,来找出图像中的部分重要内容进行后续分析处理并忽略掉冗余信息,从而加速整个任务的执行。显著性检测的结果可以应用于很多计算机视觉任务,例如物体检测和识别[9][52][80][81]、图像分割[84][85][86]、图像和视频压缩[82][83]、图像重定向[53][79][100]、视频摘要[88][89][90]、视觉跟踪[94][95][96][97][98][99]、基于内容的图像检索[87]、图像编辑[91][92][93]等。因为显著性检测的重要性,它越来越受到重视并被大量研究,很多显著性检测模型相继被提出。在计算机视觉的显著性检测领域,根据任务的不同,显著性检测有两个分支：通用的显著性检测和特定的显著性检测。而根据结果的特点,每个分支还可以进一步分为视觉显著性检测和显著性物体检测两类。通用的显著性检测任务,目的是查找自然图像中吸引人注意的区域或物体,这些区域或物体没有明确的类别。而特定的显著性检测,一般是根据不同任务在图像中查找某种类型的区域或物体,比如照片中的人脸、监控中的汽车、医学图像中的肿瘤。本文全面研究了图像内容显著性检测的理论与方法,从多个角度、多个方面分析了现有的显著性检测模型,并提出了新的特征、模型和评价方法,对显著性检测领域做出了较大贡献。主要创新点包括：总结了视觉显著性检测模型和显著性物体检测模型两类模型的发展,并且汇总了两类模型的显著性特征、评价数据库和评价方法,发现这两类模型有很多相似点。进一步分析,两类模型都有三个主要组成部分：特征对比方式、显著性提取方向以及线索结合方法,再次说明了两类模型的紧密关系。提出了一种通用的显著性物体检测模型UFO[10],在该模型中提出了聚焦度(focusness)和物体度(objectness)两种显著性特征,其中聚焦度可以通过尺度空间分析进行估计而物体度则由修改的物体检测算法计算。最后,非线性结合广泛使用的独特度(uniqueness)得到UFO模型。该模型在当时国际上最大和最难的显著性物体检测数据库MSRA1000和BSD300上,及统一的评价体系下,取得了领先的结果。提出了一种对基于扩散的显著性物体检测模型的改进方法。通过分析现有的基于扩散的显著性物体检测模型,我们对该类模型的工作机制有了一种全新的解释,我们发现基于扩散的显著性物体检测模型的性能与扩散矩阵和种子向量都有关,并且性能上限由扩散矩阵决定。因此,我们提出了一种通过重新合成扩散矩阵和构建种子向量来提高模型准确性和效率的方法。之前大多数基于扩散的显著性物体检测模型只关注于种子向量的生成,但是我们通过大量实验,包括视觉显著性提升能力、及我们提出的受限最优的种子点效率(COSE),充分证明了我们重新合成的扩散矩阵有更强的扩散能力,可以使种子向量的显著性信息更精确的传播到整个显著性物体。同时,视觉显著性提升能力的实验为改造视觉显著性检测模型来检测显著性物体提供了一个途径。最后,我们结合重新合成的扩散矩阵及构建的种子向量得到GP模型[11]。我们在当时最大的两个数据库MSRA10K和ECSSD上进行显著性物体检测的实验,GP在大多数评价方法下都取得领先水平。提出了一种特定的显著性物体检测模型。具体来说,该模型实现了一个算法,来自动的检测乳腺超声图像中的肿瘤位置并勾绘出肿瘤轮廓。该模型首先利用AdaBoost分类器找出所有潜在的肿瘤区域,再利用SVM分类器进一步把真实肿瘤区域筛选出来。最后将检测出的肿瘤区域及非肿瘤区域的中心作为前/背景种子点,利用Random Walks分割算法得到肿瘤轮廓。实验证明,该模型可以准确定位肿瘤位置并精确勾绘肿瘤轮廓,此外该算法也可以应对包含多个肿瘤的超声图像。
[Abstract]:The human vision system can quickly position the most attractive content in a large, complex dynamic and static scene, which is called the significance detection. The content of the attraction is therefore called significance (content), and there is a significant difference in some of the characteristics of the general significance, such as some dangerous warning signs. The significance detection capability of the vision system can make us more focused on a local part of the visual scene, and turn a blind eye to other background parts in the scene, so that part of the signal can be preferentially processed and the reaction can be made more quickly when we face the external stimulus. With the development of image video acquisition and capture device, the scale of the data is becoming more and more complex, so the early computer vision algorithm can't be qualified for the current task. Therefore, people try to design the algorithm to simulate the significance detection ability of the human vision system, to find out some important content in the image for subsequent analysis and to ignore the redundant information, so as to accelerate the execution of the whole task. The results of the significance detection can be applied to many computer visual tasks, such as object detection and identification[9][52][80][81], image segmentation[84][85][86], image and video compression[82][83], image redirection[53][79][100],[88][89][90], visual tracking[94][95][96][97][98][99], content-based image retrieval[87], image editing[91][92][93], and the like. Because of the significance of the significance test, it is more and more important and has been extensively studied, and many significant detection models have been proposed successively. In the field of the significance detection of computer vision, there are two branches according to the difference of the task: the general significance detection and the specific significance detection. and according to the characteristics of the result, each branch can be further divided into two types of visual significance detection and saliency object detection. The purpose of the universal significance detection task is to find areas or objects that are noticed by a person in a natural image that has no clear category. and the particular significance detection is generally used to find some type of region or object in the image according to different tasks, such as the human face in the photograph, the automobile in the monitoring, and the tumor in the medical image. In this paper, the theory and method of the significance detection of the image content are comprehensively studied, and the existing significance detection model is analyzed from a plurality of angles and a plurality of aspects, and a new characteristic, a model and an evaluation method are put forward, and a great contribution is made to the field of significance detection. The main innovation points include: the development of two models of the visual significance detection model and the significance object detection model, and the significance characteristics of the two types of models, the evaluation database and the evaluation method are summarized, and the two types of models are found to have many similar points. Further analysis, the two models have three main components: the feature contrast, the significance extraction direction and the lead-binding method, and the close relationship between the two types of models is described again. A general significance object detection model UFO[10] is proposed. In this model, two significant features of focus degree and object degree are proposed, in which the focus degree can be estimated by the scale space analysis, and the object degree is calculated by the modified object detection algorithm. Finally, the unique degree of non-linearity in combination with a wide range of uses results in a UFO model. The model has made a leading result under the unified evaluation system of the world's largest and most difficult-most significant object detection databases, MSRA1000 and BSD300, and the unified evaluation system. An improved method for detecting a significant object based on diffusion is presented. By analyzing the existing diffusion-based saliency object detection model, we have a brand-new interpretation of the working mechanism of this kind of model, and we find that the performance of the diffusion-based saliency object detection model is related to both the diffusion matrix and the seed vector. and the upper performance limit is determined by the diffusion matrix. Therefore, we propose a method to improve the accuracy and efficiency of the model by re-synthesizing the diffusion matrix and constructing the seed vector. Most of the previous diffusion-based significance object detection models focus only on the generation of seed vectors, but we have passed a number of experiments, including visual saliency enhancement, and the limited optimal seed point efficiency (COSE) we propose, It is proved that the diffusion matrix of the re-synthesis has stronger diffusion ability, and the significance information of the seed vector can be more accurately transmitted to the whole significant object. At the same time, the experiment of the visual significance enhancement capability provides a way to transform the visual saliency detection model to detect the significant object. Finally, we combine the re-synthesized diffusion matrix and the constructed seed vector to get the GP model[11]. The experiment of significant object detection on the two largest databases, MRA10K and ECSSD, was the leading level of GP under most of the evaluation methods. A particular significance object detection model is proposed. In particular, the model implements an algorithm to automatically detect the tumor location in the breast ultrasound image and to map out the tumor profile. The model first uses the AdaBoost classifier to find all potential tumor regions, and further filters the real tumor region by using the SVM classifier. and finally, the detected tumor region and the center of the non-tumor region are taken as a front/ background seed point, and a tumor profile is obtained by using the random Walks segmentation algorithm. The experimental results show that the model can accurately position the tumor position and draw the tumor contour accurately, and the algorithm can also deal with the ultrasound image containing multiple tumors.
【学位授予单位】：山东大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】