【作者】 郭明玮；

【导师】 陈宗海；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2014， 博士

【摘要】 视觉目标检测是计算机视觉中最为基础的研究方向,其任务在于从图像中检测并定位目标物体。当前,目标检测在视频监控、人机交互、智能交通、辅助驾驶、医学图像分析等领域都有着重要的应用。大多目标检测算法将目标检测的问题简化为一个二分类问题,即通过训练学习得到目标的特征模型,然后判断待检测图片中滑动窗口内是否存在目标,如果存在则定位该目标位置。所以目标检测的主要步骤包括目标模型构建、目标搜索和目标分类。由于视觉图像会受到光照变化、背景干扰、目标遮挡等方面的影响,很多目标检测算法存在鲁棒性和实时性问题。但是,人类的视觉系统则可以在更为复杂的情况下完成目标检测任务。因此,模拟人类视觉完成目标检测的机理和机制,构建具有人类视觉智能的目标检测算法,是改进目标检测算法和提高其性能的重要途径。本文深度分析了人类视觉记忆机制在完成目标检测时的工作机理,结合研究团队提出的特征学习与特征联想的概念,通过构建基于视觉记忆的特征学习与特征联想模型,开展模拟人的视觉智能的目标检测算法研究。文章的主要工作及贡献包括：(1)构建了基于视觉记忆的特征学习与特征联想模型,并在此基础上构建了基于视觉记忆的目标检测算法框架。通过研究视觉记忆在完成目标检测时的工作机理,将视觉记忆描述为一个特征学习与特征联想的过程。结合长期视觉记忆和视觉工作记忆的工作机理,采用特征处理方式和视觉记忆必要特征对视觉记忆进行了模拟。其中特征处理方式是人类视觉系统为方便快速的搜索目标而提取特征时所采用的方法,视觉记忆必要特征则是对目标进行分类判别时所需的重要特征。(2)提出了一种基于选择性注意机制的显著性检测算法来模拟视觉记忆保存的特征处理方式。为了解决显著性检测算法会受到小尺度细节图像干扰的问题,基于目标的尺度、颜色和位置信息采用了分层的显著性计算方法,从而提取输入图像的显著图。并在此基础上,通过自适应阂值对显著图进行二值化,然后从得到的二值图中提取目标检测候选区域。(3)基于特征学习与特征联想的视觉记忆模型,构建了模拟人类视觉智能的目标检测算法。该算法采用了局部可变模型作为检测器,以目标的边缘和强度特征作为主要特征来模拟视觉记忆必要特征；同时还采用了基于选择性注意机制的显著性检测算法来提取目标检测候选区域,从而缩小了目标检测的搜索区域。通过多个数据集中对本文算法的测试和分析,结果显示本文算法在保证目标检测准确率的同时提高了目标检测算法的实时性,且对于不同场景的图像均具有一定的适用性。(4)针对目前大多视频摘要生成算法无法提供目标属性信息且容易丢失目标时间和空间信息的问题,本文构建了一种基于目标检测的视频摘要生成算法。该算法基于目标检测提取目标的关键帧,并由此生成包含目标所有信息的视频摘要,同时可根据目标的属性信息来进行检索。通过该算法生成的视频摘要,人们可以在实际视频监控平台中方便快捷的浏览和搜索监控视频。(5)构建了基于目标检测的图像质量评价算法。该算法提取图像的检测器得分和显著性得分来描述图像的清晰度、背景复杂度和图像中目标的完整性,并由此对图像质量进行评价。通过多个数据集对本文图像质量算法进行测试,结果显示本文算法符合人类视觉系统评价图像质量时的特性。更多还原

【Abstract】 Object detection is one of the fundamental problems in computer vision and it focuses on detecting object from images. Object detection is widely used in many applications, such as video surveillance, human-computer interaction, intelligent traffic management, driver assistance system and medical image analysis. Today, most object detection methods prefer to simplify object detection into a binary classification problem:determine whether there is the set type of object in the sliding window or not. Thus, the main steps of an object detection method contain the construction of object model, object search strategy and the classification of object. For the reason that, image will change a lot when variable illumination, background interference or object occlusion happens. At this time, object detection method can not detect object in the image well and the detection process will be more time consuming. However, human visual system can complete the object detection well when facing the same problem. To improve the performance of object detection, it’s an important way to research the mechanisms used by human visual system when detecting object and construct an object detection algorithm witch possessed the intelligence of human visual system.On the basis of the concept of feature learning and feature imagination, this dissertation focuses on the mechanism of human visual memory and constructs the feature learning and feature imagination model based on human visual memory. Then we use this model to propose an object detection algorithm. Tested on some datasets, our object detection algorithm has been proved that it can speed up object detection without any decline in detection accuracy. On the other hand, this dissertation also discusses video synopsis and indexing method and image quality assessment method based on object detection, and analyzes their performance based on experiment. The main contents of this dissertation are listed as below:1. We study the human visual memory mechanism and describe it as a process of feature learning and feature imagination. We also describe and analyze two types of the saved feature of visual memory. The saved features of visual memory include mode of processing features and essential features for visual memory. The mode of processing features extracts visual features to make the search of object convenient, and the essential features of visual memory are the important features to classify the object. Thus we construct feature learning and feature imagination model based on visual memory.2. In order to simulate the mode of processing features in visual memory, we construct a saliency detection algorithm based on human visual selective attention mechanism. This algorithm uses scale, color and position information of object to propose a hierarchical saliency calculation method. And then, we use a binarization method based on adaptive threshold to binarize the saliency map and extract candidate detection area. Thus we reduced the search area of object detection and speed up object detection.3. Based on the feature learning and feature imagination model, we propose an object detection algorithm based on visual memory. The algorithm use deformable part model as the detector and simulate essential features of visual memory by extracting edge and intensity features of object. Our object detection algorithm can speed up object detection without any decline in detection accuracy.4. This dissertation proposes a video synopsis and indexing algorithm based on object detection. The video summary generated by our algorithm will not lose any spatial and temporal information of object, and can be retrieved by the property information of object. We analyze the performance of our algorithm based on the experiment in the video surveillance system.5. Image quality assessment is one of the hot research areas in the field of image processing. For the reason that human being is the final receiver of the image, the image quality assessment should match the characteristics of human visual system. The image quality assessment algorithm proposed in this dissertation extracts detector score and saliency score of image to describe the image clarity, complexness of background and completeness of the object in the image, and get the final image quality assessment. Tested on some datasets, our algorithm is proved to meet some characteristics of the human visual system.更多还原