Research on Terror Video Identification Technology Based on Multi-instance Learning
LIU Liping, Guangli Wu, Yang Zhilong · 2020 5th International Conference on Computer and Communication Systems (ICCCS) · 2020
The main purpose of terror video study is to analyze the emotions contained in video itself. The main content of this paper includes the following four parts: first, video in the data set was preprocessed to extract the visual features expressing horrible video, such as optical flow features, color moment and other traditional features. In addition, relevant features were extracted by using convolutional neural network. Second, the features of each frame in video are regarded as instances, and video is expressed as positive and negative packages of multi-instance learning. Then select visual semantics to construct visual projection space according to the principle of maximum point density; Then the nonlinear function is used to calculate the feature of visual projection. Thirdly, support vector machine and extreme learning machine training classifier are used to classify and identify whether it is terror video or not. Fourthly, relevant experiments are carried out to analyze the differences between traditional features and features extracted by convolutional neural network as well as two algorithms, support vector machine and extreme learning machine, and to evaluate the model.