人眼关注点驱动的视觉显著性检测算法研究
发布时间：2018-07-18 15:11
【摘要】：随着互联网技术的不断发展,计算机需要处理海量的数据,而在这海量的数据中,80%的信息是视觉信息,因此快速准确的处理视觉信息,完成图像或视频的分析任务是一个研究的热点。本文受人眼关注点所驱动进行视觉显著性检测算法的研究。依据人类视觉机制可知,人类在观察图像或视频时,会将注意力集中在某些人眼感兴趣的区域,而忽略相对不显著的区域。视觉显著性检测的主要任务是提取图像或视频中的显著区域,尽可能准确的模拟人眼关注区域。视觉显著性检测研究主要分为两个方向,一是显著物体检测,二是预测人眼关注点。本文的主要研究方向是建立预测人眼关注点的视觉显著性检测模型。本文基于对现有显著性检测算法的研究,提出了基于重启型随机游走模型的视频显著性检测算法,该算法首先利用SLIC超像素分割算法将视频的每一帧分割成超像素构建无向图,然后提取视频中YUV颜色空间的空域特征建立空域转移概率矩阵,提取H.264码流信息中的运动矢量用随机游走模型建立时域重启矩阵,再利用边界先验基于吸收马尔科夫链建立背景先验的重启矩阵,最后用改进的重启型随机游走模型迭代计算得到最终的显著性检测结果图。该算法在公开的SFU和CRCNS视频集上进行了仿真实验,实验结果表明,本文提出的基于重启型随机游走模型的视频显著性检测算法不仅具有较低的时间复杂度,而且有较高的检测准确度,有效的抑制了背景噪声。由于目前基于压缩域的显著性检测逐渐受到关注,本文同时提出了一种基于互信息的压缩域视频显著性检测算法,该算法首先从H.264压缩码流中提取空域特征和时域特征,利用香农互信息的基本理论分析显著性,通过计算每个宏块的中心窗和邻域窗内的特征集间的互信息确定空域显著图和时域显著图,然后自适应融合得到时空域显著图,再用基于凸包的中心先验优化时空域显著图获得最终的显著图,有效地提高了检测的准确度。同样,对算法在公开的视频集上进行仿真,实验进一步验证了算法的有效性。
[Abstract]:With the development of Internet technology, computers need to deal with huge amounts of data, in which 80% of the information is visual information, so the rapid and accurate processing of visual information, The task of image or video analysis is a hot topic. In this paper, visual salience detection algorithm driven by human eye concern is studied. According to the mechanism of human vision, when we observe images or videos, we will focus our attention on the regions of interest to some people's eyes, while ignoring the relatively insignificant areas. The main task of visual salience detection is to extract salient regions from images or videos and to simulate human eye regions as accurately as possible. Visual salience detection is mainly divided into two directions, one is significant object detection, the other is predicting human eye concern. The main research direction of this paper is to establish visual salience detection model to predict human eye concern. Based on the research of the existing salience detection algorithms, this paper proposes a video saliency detection algorithm based on the restart random walk model. Firstly, the SLIC super-pixel segmentation algorithm is used to segment every frame of the video into super-pixels to construct undirected graph. Then the spatial feature of YUV color space in video is extracted to establish the spatial transfer probability matrix, and the motion vector of H.264 bitstream information is extracted to establish the time domain restart matrix using random walk model. Then the background priori reboot matrix is established based on the absorbing Markov chain and the final significance detection result is obtained by iterative calculation of the improved restart random walk model. The simulation results on SFU and CRCNS video sets show that the proposed video saliency detection algorithm based on restart random walk model has low time complexity. Moreover, it has high detection accuracy and effectively suppresses background noise. Because the salience detection based on compressed domain has been paid more and more attention to, this paper also proposes a video saliency detection algorithm based on mutual information. Firstly, this algorithm extracts spatial and temporal features from H.264 compressed bitstream. Based on the Shannon mutual information theory, the spatial and temporal salience maps are determined by calculating the mutual information between the center window and neighborhood window of each macroblock, and then the spatio-temporal significant map is obtained by adaptive fusion. Finally, the final salience map is obtained by using the center prior optimization of saliency map based on convex hull, and the accuracy of detection is improved effectively. In the same way, the algorithm is simulated on the open video set, and the effectiveness of the algorithm is further verified by experiments.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
本文编号：2132298
[Abstract]:With the development of Internet technology, computers need to deal with huge amounts of data, in which 80% of the information is visual information, so the rapid and accurate processing of visual information, The task of image or video analysis is a hot topic. In this paper, visual salience detection algorithm driven by human eye concern is studied. According to the mechanism of human vision, when we observe images or videos, we will focus our attention on the regions of interest to some people's eyes, while ignoring the relatively insignificant areas. The main task of visual salience detection is to extract salient regions from images or videos and to simulate human eye regions as accurately as possible. Visual salience detection is mainly divided into two directions, one is significant object detection, the other is predicting human eye concern. The main research direction of this paper is to establish visual salience detection model to predict human eye concern. Based on the research of the existing salience detection algorithms, this paper proposes a video saliency detection algorithm based on the restart random walk model. Firstly, the SLIC super-pixel segmentation algorithm is used to segment every frame of the video into super-pixels to construct undirected graph. Then the spatial feature of YUV color space in video is extracted to establish the spatial transfer probability matrix, and the motion vector of H.264 bitstream information is extracted to establish the time domain restart matrix using random walk model. Then the background priori reboot matrix is established based on the absorbing Markov chain and the final significance detection result is obtained by iterative calculation of the improved restart random walk model. The simulation results on SFU and CRCNS video sets show that the proposed video saliency detection algorithm based on restart random walk model has low time complexity. Moreover, it has high detection accuracy and effectively suppresses background noise. Because the salience detection based on compressed domain has been paid more and more attention to, this paper also proposes a video saliency detection algorithm based on mutual information. Firstly, this algorithm extracts spatial and temporal features from H.264 compressed bitstream. Based on the Shannon mutual information theory, the spatial and temporal salience maps are determined by calculating the mutual information between the center window and neighborhood window of each macroblock, and then the spatio-temporal significant map is obtained by adaptive fusion. Finally, the final salience map is obtained by using the center prior optimization of saliency map based on convex hull, and the accuracy of detection is improved effectively. In the same way, the algorithm is simulated on the open video set, and the effectiveness of the algorithm is further verified by experiments.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
【参考文献】
相关期刊论文 前2条
1 吴则举,陈俊东,刘云,Roemer Louis;静止背景的视频对象分割[J];青岛科技大学学报(自然科学版);2004年05期
2 杨威,张田文;复杂景物环境下运动目标检测的新方法[J];计算机研究与发展;1998年08期
相关博士学位论文 前2条
1 景慧昀;视觉显著性检测关键技术研究[D];哈尔滨工业大学;2014年
2 单列;视觉注意机制的若干关键技术及应用研究[D];中国科学技术大学;2008年
相关硕士学位论文 前1条
1 姜博文;基于马尔可夫链的显著性检测[D];大连理工大学;2014年
，本文编号：2132298
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2132298.html
