Research And Application Of Concept Detection In Video Retrieval

With the rapid development of internet technology, the popularity of digital camera equipment and storage capacity increase, the amount of video data shows an explosive growth. How to conduct effective mass video data retrieval has become an urgent demand. At present, most of Internet's video search engines use video metadata, such as title, profile and so on, through the text retrieval approach to realize video retrieval. The shortcoming of this search approach is obvious: it is entirely dependent on man-made descriptive text, which is often too simplistic and subjective, and the massive video data annotation is very time-consuming and labor-intensive. In recent years, semantic content-oriented video retrieval attracts more and more attention, and its aim is to achieve the automatic semantic video analysis and retrieval. In semantic content-oriented video retrieval, the concept detection is the most critical technology. Its purpose is to automatically detect the video contains a large number of basic semantic concepts, such as cars, sand, sky and so on. When a large number of semantic concept descriptors were build up, semantic concepts based video retrieval would be the effective way to achieve content-oriented video retrieval. Therefore, the research of concept detection is of great significance for video retrieval development and application.This article focuses on the following two kinds of algorithm and one demo system.First, concept detection is considered a machine learning problem, including feature extraction and supervised learning. This paper presents a validation average precision based multi-feature late fusion method. It effectively improves the search results, and has a good robustness.Secondly, for the co-existence relationships between semantic concepts, taking two concepts as examples, "female face" and "human face", this article proposes a method of using related concepts to carry out pre-filtering and post-filtering, and analyses the fault-tolerant performance of this method. The experimental results show that the method can effectively improve the related concept detection average precision. Lastly, based on the study of video retrieval systems of the current majorinternational research institutions, this article designed a semantic concept based video retrieval system. Experiments show that the mean average precision of the system is pretty good.