Research Of Web Video Quality Assessment And Related Technology

With the prevailing of video sharing websites, more and more people tend to upload videos for sharing with other people. Meanwhile, taking videos becomes a trend in public and people can taking their own videos anytime, anywhere in any situations. With the information explosion, in contemporary, this kind of behavior brings abundant multimedia information quantity for Internet. However, in another hand, the quality of the videos provided by search engines is more and more difficult to control, because of the various videos with different quality uploaded by customers. Hence, it is very necessary to design a set of quality assessment algorithm for videos on Internet.To acquire the target mentioned above, in this paper, a set of web video assessment algorithm based on pattern recognition technology, or called for web video assessment system. In this algorithm, the input is a segment of web video and the output is the quality score or label of this segment of web video.There are four main parts in the algorithm realization:1. Design and implement of the pre-processing method for the input videos. Specifically, the pre-processing consists of video segmentation, key frame and key shot extraction and the image processing of exception detection, black side removal and normalization.2. Design and implement features utilized for web video quality assessment which includes three aspects. They are spatial feature based on information in frames, temporal feature based on global motion estimation and structure feature based on the shot scale.3. Train and test the performance of the designed features with support vector machine(SVM) and Adaboost classifier. In this paper, the author adopts the famous LibSVM library for SVM classifier and realizes the Adaboost classifier by his own hand.4. Gather and design a dataset for testing the performance of video quality assessment algorithm. In this dataset, there are2498labeled videos and the total length is more than122hours. Then the experiment is based on SVM and Adaboost classifiers. According to the experiment result, the merit and deficiency and its reason are analyzed. In the end, some comments and advises are given.