视觉显著性检测模型研究及应用
发布时间：2018-08-02 10:46
【摘要】：人类视觉系统在面对复杂自然场景时,具有快速搜索感兴趣目标的能力,这种能力我们称之为视觉注意。在人类生存与发展的过程中,视觉注意扮演着至关重要的角色。视觉注意和人类如何感知、处理视觉刺激紧密相关,并且正在被包括认知心理学、神经生物学和计算机视觉在内的多个学科进行研究。随着认知心理学和神经生物学的不断发展,通过对视觉机理的研究发现,人类视觉对场景中目标的选择性可分为两个阶段：一个快速的、无意识的、数据驱动的、自底向上的阶段和一个较慢的、有意识的、任务驱动的、自顶向下的阶段.而与视觉注意紧密相连的概念就是视觉显著性,他是指导视觉注意的一个关键注意机制。图像显著性区域检测研究的目的是快速定位显著性区域并反映显著性区域的显著程度。视觉显著性区域检测在图像处理中有着广泛的应用,包括图像分割、目标识别、自适应压缩、内容敏感图像编辑、图像检索、目标检测、目标跟踪、图像质量评价等。本文从视觉注意机制的研究出发,对视觉显著性检测与应用中的一些关键问题进行了较为深入的研究,提出了一些新的思想和算法。论文的主要工作与贡献包括：(1)针对已有局部对比度和全局对比度建模方法存在的不足,本文提出了一种基于条件随机场融合全局特征的显著性区域检测方法。该方法首先采用唯一性、颜色空间分布等全局特征计算相应的显著图：其次在条件随机场框架下融合多个显著图,通过显著性区域与背景区域的区域标注实现显著性区域初步检测；然后采用基于显著性区域的高斯模型计算目标先验图,并对全局特征显著图进行高斯滤波；最后再利用条件随机场融合滤波之后的显著图来实现更加精确的显著性检测。实验结果表明该方法能均匀致密的凸显显著性区域,有效的抑制背景干扰,并具有较高的检测准确率与召回率。(2)基于视觉机制挖掘可应用的更高层次的显著性先验特征,本文提出了一种融合多级显著性特征的显著性目标检测方法。该方法融合了基于像素级的局部对比度、基于区域级的全局对比度以及基于目标级的背景先验信息。该方法基于凸包检测技术使用底层的视觉线索从背景分离显著性目标。基于初级的检测结果提取背景模版,利用PCA计算背景先验信息。为了抑制背景干扰,该方法采用目标中心先验信息精炼局部对比度特征和全局对比度特征。在公开的数据集上的实验表明,该方法所得到的显著图能较好的凸显显著性目标.同时也证明Otsu自适应阈值方法可以用来产生高质量的目标分割结果。(3)针对视觉显著性在目标跟踪过程中的应用研究,本文提出了一种基于视觉注意的目标跟踪算法。该算法首先采用基于背景先验的视觉显著性检测算法来提取目标的显著性特征,其次采用基于贝叶斯决策理论的前景背景分类方法来提取目标的运动特征,然后利用显著特征引导运动特征与颜色特征进行目标状态估计,最后结合自适应粒子滤波形成目标跟踪算法。实验结果表明在复杂场景下,该算法相对于现有的目标跟踪算法具有较强的鲁棒性,对光照变化、姿态变化、目标遮挡、快速运动、复杂背景等具有较好的跟踪效果。(4)针对枪球联动接力跟踪过程中的目标离开枪机画面后在球机中初始定位问题,本文提出了一种基于视觉注意的枪球联动接力跟踪方法。该方法采用网格结合插值算法实现在球机中的目标放大跟踪,当目标离开枪机画面时,利用视觉显著性检测算法计算候选区域,利用枪机保存的目标模版在候选区域中搜索匹配区域,确定目标在球机场景中的位置,最后利用Mean Shift跟踪算法实现球机的主动跟踪。实验结果表明本文提出的基于视觉注意的枪球联动接力跟踪具有较好的实时跟踪效果。
[Abstract]:In the face of complex natural scenes, the human visual system has the ability to quickly search for a target of interest, which we call visual attention. Visual attention plays a vital role in the process of human survival and development. Visual attention and human perception are closely related to visual stimuli, and are being included. The study of cognitive psychology, neurobiology and computer vision. With the continuous development of cognitive psychology and neurobiology, the study of visual mechanisms found that human vision can be divided into two stages: a fast, unconscious, data driven, bottom-up. The stage and a slow, conscious, task driven, top-down stage. The concept of close connection with visual attention is visual significance. He is a key attention mechanism to guide visual attention. The purpose of the image saliency region detection study is to quickly locate the significant region and reflect the significance of the significant region. Degree. Visual saliency region detection has a wide range of applications in image processing, including image segmentation, target recognition, adaptive compression, content sensitive image editing, image retrieval, target detection, target tracking, image quality evaluation and so on. This paper, starting from the research of visual attention mechanism, discusses some key points in visual significance detection and application. Some new ideas and algorithms are proposed. The main work and contributions of this paper are as follows: (1) in view of the shortcomings of the existing local contrast and the global contrast modeling method, a significant regional detection method based on the global characteristics of conditional random fields is proposed. We use the uniqueness, the color space distribution and other global characteristics to calculate the corresponding significant graphs. Secondly, multiple significant graphs are fused under the conditional random field framework, and the significant region detection is realized through the regional annotation of the significant region and the background region. Then the Gauss model based on the saliency region is used to calculate the prior map of the target, and the whole area is calculated. The characteristic salient image of the bureau is used to carry out Gauss filtering; finally, a more accurate detection is achieved by using the salient graph following the fusion filter of the airport. The experimental results show that the method can highlight the significant region, effectively suppress the background interference, and have high detection accuracy and recall. (2) the visual machine is based on the visual machine. This method combines the local contrast based on the pixel level, the global contrast based on the regional level and the backview prior information based on the target level. This method is based on the convex packet detection technique. The underlying visual cues are used to separate the significant targets from the background. The background template is extracted based on the primary detection results and the background information is calculated using the PCA. In order to suppress the background interference, the method uses the target center prior information to refine the local contrast characteristics and the global contrast characteristics. The experiment on the open data set shows that this method is used to extract the background information. The remarkable graph obtained by the method can better highlight the significant target. It also proves that the Otsu adaptive threshold method can be used to produce high quality target segmentation results. (3) aiming at the application of visual significance in target tracking, a target tracking algorithm based on visual attention is proposed in this paper. The visual saliency detection algorithm of background prior is used to extract the significant feature of the target. Secondly, the foreground background classification method based on Bayesian decision theory is used to extract the motion features of the target, and then the target state is estimated by using the significant feature to guide the motion feature and color feature. Finally, the adaptive particle filter is combined to form the target state. The experimental results show that in the complex scene, the algorithm has strong robustness against the existing target tracking algorithm, and has good tracking effect on illumination change, attitude change, target occlusion, fast motion, complex background and so on. (4) in the course of the gun ball linkage relay tracking, the target leaves the gun frame. In this paper, the problem of initial positioning in the ball machine is presented. This paper proposes a method of tracking the joint force of the gun ball based on visual attention. This method uses the mesh and interpolation algorithm to achieve the target amplification and tracking in the ball machine. When the target leaves the gun, the candidate region is calculated by the visual significance detection algorithm, and the target template saved by the gun is used. In the candidate region, the matching area is searched, the location of the target in the ball scene is determined, and the Mean Shift tracking algorithm is used to achieve the active tracking of the ball machine. The experimental results show that the tracking of the gun ball joint relay based on visual attention has good real-time tracking effect.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41
，
本文编号：2159180
[Abstract]:In the face of complex natural scenes, the human visual system has the ability to quickly search for a target of interest, which we call visual attention. Visual attention plays a vital role in the process of human survival and development. Visual attention and human perception are closely related to visual stimuli, and are being included. The study of cognitive psychology, neurobiology and computer vision. With the continuous development of cognitive psychology and neurobiology, the study of visual mechanisms found that human vision can be divided into two stages: a fast, unconscious, data driven, bottom-up. The stage and a slow, conscious, task driven, top-down stage. The concept of close connection with visual attention is visual significance. He is a key attention mechanism to guide visual attention. The purpose of the image saliency region detection study is to quickly locate the significant region and reflect the significance of the significant region. Degree. Visual saliency region detection has a wide range of applications in image processing, including image segmentation, target recognition, adaptive compression, content sensitive image editing, image retrieval, target detection, target tracking, image quality evaluation and so on. This paper, starting from the research of visual attention mechanism, discusses some key points in visual significance detection and application. Some new ideas and algorithms are proposed. The main work and contributions of this paper are as follows: (1) in view of the shortcomings of the existing local contrast and the global contrast modeling method, a significant regional detection method based on the global characteristics of conditional random fields is proposed. We use the uniqueness, the color space distribution and other global characteristics to calculate the corresponding significant graphs. Secondly, multiple significant graphs are fused under the conditional random field framework, and the significant region detection is realized through the regional annotation of the significant region and the background region. Then the Gauss model based on the saliency region is used to calculate the prior map of the target, and the whole area is calculated. The characteristic salient image of the bureau is used to carry out Gauss filtering; finally, a more accurate detection is achieved by using the salient graph following the fusion filter of the airport. The experimental results show that the method can highlight the significant region, effectively suppress the background interference, and have high detection accuracy and recall. (2) the visual machine is based on the visual machine. This method combines the local contrast based on the pixel level, the global contrast based on the regional level and the backview prior information based on the target level. This method is based on the convex packet detection technique. The underlying visual cues are used to separate the significant targets from the background. The background template is extracted based on the primary detection results and the background information is calculated using the PCA. In order to suppress the background interference, the method uses the target center prior information to refine the local contrast characteristics and the global contrast characteristics. The experiment on the open data set shows that this method is used to extract the background information. The remarkable graph obtained by the method can better highlight the significant target. It also proves that the Otsu adaptive threshold method can be used to produce high quality target segmentation results. (3) aiming at the application of visual significance in target tracking, a target tracking algorithm based on visual attention is proposed in this paper. The visual saliency detection algorithm of background prior is used to extract the significant feature of the target. Secondly, the foreground background classification method based on Bayesian decision theory is used to extract the motion features of the target, and then the target state is estimated by using the significant feature to guide the motion feature and color feature. Finally, the adaptive particle filter is combined to form the target state. The experimental results show that in the complex scene, the algorithm has strong robustness against the existing target tracking algorithm, and has good tracking effect on illumination change, attitude change, target occlusion, fast motion, complex background and so on. (4) in the course of the gun ball linkage relay tracking, the target leaves the gun frame. In this paper, the problem of initial positioning in the ball machine is presented. This paper proposes a method of tracking the joint force of the gun ball based on visual attention. This method uses the mesh and interpolation algorithm to achieve the target amplification and tracking in the ball machine. When the target leaves the gun, the candidate region is calculated by the visual significance detection algorithm, and the target template saved by the gun is used. In the candidate region, the matching area is searched, the location of the target in the ball scene is determined, and the Mean Shift tracking algorithm is used to achieve the active tracking of the ball machine. The experimental results show that the tracking of the gun ball joint relay based on visual attention has good real-time tracking effect.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41
，
本文编号：2159180
本文链接：https://www.wllwen.com/shoufeilunwen/xxkjbs/2159180.html
最近更新
教材专著
