Violent Video Classification Based on Spatial-Temporal Cues Using Deep Learning
Xingyu Xu, Xiaoyu Wu, Ge Wang, Huimin Wang · 2018
The rapid development of Internet technology brings convenience to our life and also brings various hidden dangers. Violent video is one of the hidden dangers. Therefore, this paper proposes a P3D-LSTM recognition method based on multi-feature fusion for violent video recognition. In this paper, starting from video's static image, frame difference image and optical flow feature, the neural network for extracting corresponding features is constructed respectively, and then late fusion method is adopted to fuse the features or decision scores to obtain video classification labels. Finally, the experiment is carried out on two public databases and self-built violent database. As far as the recognition accuracy is concerned, this method has certain application prospect in classify violent video.