基于内容的视频检索与分类方法研究

发布时间：2018-02-23 19:08

本文关键词： 视频检索镜头分割关键帧提取视频分类　出处：《沈阳大学》2013年硕士论文　论文类型：学位论文

【摘要】：随着时代的进步与科技的发展，浩如烟海的视频数据表现了社会与生活的方方面面。如何对视频信息进行检索与分类，当前已经成为一个迫切需要解决的课题。为了有效地从视频媒体库中获得所需要的信息，必须对视频信息进行有效地组织与索引。因此，基于内容的视频检索与分类方法研究符合社会与人们的需求。本文针对基于内容的视频检索与分类技术中的视频数据特点、视频结构化以及关键技术等做出了概要论述，并着重研究了视频检索与分类中的镜头分割与关键帧提方法。在现有的视频分割研究成果基础上,提出了一种基于自适应双阈值的改进算法。该算法采用权重不同的优化分块策略，并通过剔除一部分影响较大的帧间差值来减少了外围因素的干扰,与改进前算法比较，突变阈值稍有降低，提高了镜头边界检测的查全率，同时渐变阈值也稍有降低，克服了渐变中帧间差别很小的帧的影响。本文还在前人视频关键帧提取的研究基础上,提出了一种改进基于互信息的视频关键帧提取算法。该算法对关键帧数目的确定进行了优化，，使关键帧数目能够根据视频内容自动调整大小，增加了关键帧数目的自适应性，并且将以前单个镜头的关键帧提取扩展到了多个镜头以至于整个视频的关键帧提取，最终使提取的关键帧更好的描述视频内容。同时，本文用主成分分析对提取的特征进行降维处理，利用遗传算法来达到SVM分类器参数优化的目的。在保证识别精度的前提下减小特征维数，对颜色特征进行优化，找出更有利于准确分类的特征子集。同时，优化分类器的参数选取来提高分类器的分类准确率和分类速度。本文还对视频检索与分类系统进行模块化设计，并分别详细的介绍了各个模块。实验结果表明，在测试集上,本文的镜头边界检测改进算法的平均查全率和查准率均高于自适应双阈值算法，分别达到了87.86%和93.91%,取得了很好的镜头边界检测效果。本文的关键帧提取改进算法针对总帧数为12470，镜头数为115的动漫视频进行关键帧提取，提取了21个视频帧为关键帧，而未改进算法针对总帧数为605，镜头数为11的动漫视频进行关键帧提取，提取了13帧作为关键帧，上述数据表明，本算法提取的关键帧，可以有效地概括视频的内容，并且提高提高了提取效率，减少了一定的关键帧冗余。本文最后对研究工作进行总结,提出下一步工作的努力方向。
[Abstract]:With the progress of the times and the development of science and technology, the vast amount of video data shows all aspects of society and life. How to retrieve and classify video information, In order to obtain the needed information from video media library effectively, it is necessary to organize and index the video information effectively. Content-based video retrieval and classification methods meet the needs of society and people. In this paper, the characteristics of video data in content-based video retrieval and classification technology, video structure and key technologies are briefly discussed. The methods of shot segmentation and key frame extraction in video retrieval and classification are studied. An improved algorithm based on adaptive double threshold is proposed. The algorithm adopts optimized block strategy with different weights, and reduces the interference of peripheral factors by eliminating some significant inter-frame differences, compared with the improved algorithm. The abrupt threshold is reduced slightly, the recall rate of shot boundary detection is improved, and the gradient threshold is reduced slightly, which overcomes the influence of the frame with little difference between frames. This paper also based on the research of key frame extraction in previous video. An improved video key-frame extraction algorithm based on mutual information is proposed, in which the key frame number is optimized, and the key frame number can be automatically resized according to the video content, thus increasing the self-adaptability of the key frame number. And the key frame extraction of the previous single shot is extended to multiple shots so that the key frame of the whole video can be extracted so that the extracted key frame can describe the video content better. At the same time, In this paper, principal component analysis (PCA) is used to reduce the dimension of the extracted features, and genetic algorithm is used to optimize the parameters of the SVM classifier. Under the premise of ensuring the recognition accuracy, the feature dimension is reduced, and the color features are optimized. At the same time, the parameter selection of classifier is optimized to improve the classification accuracy and classification speed. This paper also designs the video retrieval and classification system modularized. Each module is introduced in detail. The experimental results show that the average recall and precision of the improved shot boundary detection algorithm are higher than that of the adaptive double threshold algorithm in the test set. In this paper, the improved key frame extraction algorithm is used to extract key frames for animation video with 12470 frames and 115 shots, and 21 video frames are extracted as key frames. However, the unimproved algorithm extracts the key frames of the animation video with 605 frames and 11 shots, and extracts 13 frames as the key frames. The above data show that the key frames extracted by this algorithm can effectively generalize the content of the video. It also improves the extraction efficiency and reduces the key frame redundancy. Finally, this paper summarizes the research work and puts forward the direction of the next work.
【学位授予单位】：沈阳大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【相似文献】