Research Of Video Shot Boundary Detection

Dividing a video sequence into shots is the first step towards video-content analysis and content-based video browsing and retrieval. This thesis addresses the key technical problems arsing in shot boundary detection and video caption detection, proposes the new methodologies and algorithms for them.This thesis synopsizes and analyzes the most methods and evaluation metrics for shot boundary detection. Expatiates on Color histograms-based methods, a new two-Gaussian model based technique is proposed that is a betterment of twin-comparison method and can detect cuts and gradual transitions, experiment demonstrates it enable to locate boundaries exactly and quickly.Puts forward a novel method based on gradient angle-module (AM) histogram, it is insentive to motion and flashlight, furthermore, we improve AM histogram and obtain blockwise AM histogram, one-dimensional histogram with comparising their performance, statistical experiments results argue that gradient information as features should be more excellent than HSV histogram, and be the same with the gradual transitions. First time try to adopt Kohonen neural network to cluster shot boundary, thus the influence of the parameters on the detection performance can be diminished and detection can be made more robust to the parameters, but it cost more time.The recent research works about video caption detection are summarized detailedly, and compared with each other. Based on segmenting video by AM histogram, this thesis proposes a new method to detect video caption by making the best of the features of video and caption, experiments show that it can locate caption exactly and quickly.