Research On Corner Detection Technique Application In News Video Retrieval

With the fast development of multimedia technologies, video has become an important information carrier and video retrieval has truned to be a new requirement for many applications. The majority of current video retrieval methods still stay in the keyword-based retrieval area, and that kind of retrieval requires the keywords to be manually pre-defined, which leads to a waste of manpower and also suffers from subjectivity problem. Consequently, content-based video retrieval has emerged as the new concerns for computer science researchers. Among the various types of videos, news video acts as an important information media platform that facilitates people to know all the events happening around the world. Therefore, people need novel and personalized techniques for news video retrieval. Meanwhile, although videos are difficult to describe due to their unstructured characteristic, news videos have their special intrinsic features, which make the retrieval of news videos become a hot topic in research area.In response to the above-mentioned requirements, this thesis studies on the application of corner detection techniques for news video retrieval based on "retaining the important features of the image from corners as well as effectively reducing the amount of data information" characteristics. Firstly, this thesis adds the information about the number of corner points into the shot boundary detection algorithm. The proposed algorithm is based on color-histogram in order to prevent the undetected errors which are caused by similar color-histograms structure but different contents. Secondly, this thesis proposes a new method, which utilizes the technique of corner detection to identify the subject caption frames in gray-scale color space. Then the video abstraction is set up by extracting reasonable caption frames under certain regulations. Finally, the exact location of the text area is got through clustering the corners by DBSCAN. and image processing techniques are applied to the subtitles to enhance the performance of the OCR recognition.As is shown in our experiments, firstly, our improved shot segmentation algorithm reduces the missed errors to a certain extent regarding the sequence shots which have the similar color-histograms but the different contents. Secondly, the video summary of the news video which is extracted by the proposed algorithm can summarize each section of the video primely, and the number of the extracted key frames is quite reasonable. Finally, the algorithm proposed by this thesis is better than text area extraction algorithm based on edge detection in performance because of the different detecting granularities. The subtitles OCR recognition results are better after processed by the proposed method in this thesis.