【作者】 于哲舟；

【导师】 周春光；

【作者基本信息】 吉林大学 ， 计算机应用技术， 2007， 博士

【摘要】 视频流图像内容检索与运动目标检测研究是在监控信息流中融入图像理解技术。图像理解是关于图像内容的描述或者是某种分类,其描述或分类是由计算机代替人的视觉完成,并由计算机自动给出对图像的评价。根据捕获到的图像对实际物体和场景做出有意义的判定,并对其行为进行理解与描述,实现“视觉智能”。本文针对图像颜色分布特征,提出对图像进行子块分割和合理的区域划分方法来对不同子块和区域分别进行颜色特征提取,并给出利用离散余弦变换和奇异值分解进行颜色特征提取和降维操作算法。提出自相似特征编码方法,将编码空间由灰度空间扩展到彩色空间,充分体现了图像的颜色自相似特征。采用中心扩散算法,能够在保证一定程度匹配误差的基础上有效缩短计算时间,增强算法的可行性,通过对自相似特征编码进行奇异值分解的方法提取特征向量来实现图像检索。文中给出利用提升小波变换和自适应二进制算术编码方法的图像压缩算法,提出动态压缩倍数方法,较好地解决了网络带宽问题。文中最后提出一种嵌入式系统工程化设计思想,并依此设计了一种网络摄像机硬件架构。讨论了嵌入式硬件实现中的关键问题-算法并行性、数据复用性、系统安全性问题。更多还原

【Abstract】 The purpose of image understanding lies in intelligentizing image information processing, moreover it can replace parts of human work. Image understanding is a kind of description or certain sort concerning image contents, its description or sort is completed and evaluated by the computer instead of human vision. The object of computer vision research is to make meaningful judgment on practical objects and scenes according to image construction scene and acquired images, comprehend and describe its behavior and realize“vision intelligence”The research of this paper is based on the leading projects“Identity authentication system based on biologic feature recognition”and“Key science and technology development plan project of Jilin province”of the national 863 plan, and is applied in the image retrieval and motion object detection of network intelligent camera. By performing motion detection to the images in video stream and retrieval based on content on sensitive information, we can make judgment to the given images, such as, identify“human”activities in surveillant scene, detect“fire”alarm information in the forest fire monitoring situation etc. This paper focuses on the vision information of images and discusses the problems of retrieval based on content and object detection.Color feature is not neglectable in the process of image retrieval because it’s very straight forward. This paper focuses on the color distribution feature of image, and puts forward to extract color feature of different sub blocks and regions respectively by sub block partition to the image and reasonable region, and introduces an algorithm that makes use of DCT and SVD to extract color feature and reduce dimension. The experiment proves that the method has achieved a better complete check rate and accuracy check rate, having preferable retrieval effect. At the same time we analyze the limitation of color feature and discuss method that combines texture, shape and fractal, trying to achieve better retrieval effect. In order to fully express the self resemblance color feature of the images, this paper puts forward the self resemblance feature coding method, which expands the code space from the gradation dimension to chromatic dimension. At the same time, central diffusion algorithm which can effectively shorten computing time, enhance the feasibility of the algorithm while ensuring certain degree of matching error is adopted because of the great computation quantity, low speed problems of the traditional fractal encoding. Finally, image retrieval can be implemented by extracting feature vector from the self resemblance feature coding using singularity value decomposition. The experimental result shows that this algorithm has better real-time quality.The given image compression algorithm makes use of lifting wavelet transformation and self-adapting binary arithmetic coding method, and presents the dynamic variety diploid method which solves the network bandwidth problem preferably.At last this paper introduces an embedded system engineering design idea and designs the hardware architecture of network camera based on it. We discuss the key problems in the embedded hardware implementation-parallelism of algorithm, data reusing, and system security. We resolve the huge amount of information and limited bandwidth resource problems that as merely an input end, the traditional network camera will send the image information captured directly to the server without any processing 24 hours ceaselessly. At the same time, this paper presents a retrieval method that can detect accidents and security hidden trouble immediately from massive information. Security monitoring has experienced three generations of development, the first generation is simulated image monitoring, the second is monitoring based on PC technology, the third is network digital monitoring. The“network camera based on image retrieval”which has“vision intelligence”and based on the research achievement of this paper belongs to the new generation network monitoring machine.“Image content retrieval and motion object detection of video stream”is a problem that involves multiple subjects. The research of image retrieval and motion object detection have obtained some achievement, but compared to human perception ability, there is still a large gap between them. The successive research of this paper will try to apply semantic technique in the compile technology in the research of image understanding and plan to solve the difficulty of how to extract semantic feature from images and describe it. Searching and developing new technology can help to improve image target identification and understanding, it also can reduce the computing complexity effectively.更多还原