Hadoop Based Video Processing And Retrieval System

As digital cameras and the social network become more and more prevalent, billions of videos are emerging throughout the Internet, and so comes the problem of video data management. Compared with words pictures or sounds, videos are easier to produce, contain much more information, more vivid and attractive, thus playing a significant role while people sharing their life and feelings. Meanwhile, video files are not as format as text files and always too large, making great trouble when coming to organizing and managing.The prime work for video management is video retrieval. The biggest challenge for video retrieval is how to extract high level semantics from the unstructured, space and time exist side by side, and large binary data. The traditional method is text based video retrieval, which use video tags to build index and search videos. But the raw videos do not carry any tag, and the small amount which with tags are not good enough to be used directly, what’s more, Human-Driven Labeling costs too much time and money. In one word, text based video retrieval cannot meet our demand. Another effective method is Content Based Video Indexing and Retrieval (CBVIR), a kind of Machine-Driven Labeling method. CBVIR is based on the information or features of multimedia data to retrieve information effectively. The information or features refers to frames shots scene of videos, or color texture framework and object from images, tone tune and melody in the audio data. As for CBVIR, we should make a reasonable model first, then do data partition and feature extraction, then do video retrieval on this model. Techniques based on CBVIR are widely used in many areas, for example remote monitoring, multimedia conference, virtual reality, news on TV, entertainment videos, and even weather analysis and prediction system, making great changes in multimedia industry.In this paper, we developed a video processing and retrieval system based on CBVIR. We first use some video processing tools to analyze video contents and present these contents in text format, then build the video index based on these data, and finally we develop a Web search engine for the user to search video files. This system extract visual and audio data from video data, and then do automated image annotation and automated speech recognition separately on these data, at last combine these two channel information and build video index. For the Web search engine part, we combine Ajax, JSP and Servlet techniques, user input their queries in the browser, and the server extract the keyword from the query, then do search in the video index, and last give the results back to the browser.We also do video data processing on Hadoop platform, which is a new idea in video retrieval work, in this way we can dispense tasks on different compute nodes easily, making the full use of compute resources, thus saved time and improved processing efficiency. By providing a friendly user interface, the video search results can be seen directly, and we also accomplished the one-click processing procedure in our UI through user identification. Our system also support video database updating, anytime you could add some new videos into this system and finish data processing in a short time and then retrieval these videos quickly.