A Deep Learning based Approach for Precise Video Tagging

摘要

With the increase in smart devices and abundance of video contents, efficient techniques for the indexing, analysis and retrieval of videos are becoming more and more desirable. Improved indexing and automated analysis of millions of videos could be accomplished by getting videos tagged automatically. A lot of existing methods fail to precisely tag videos because of their lack of ability to capture the video context. The context in a video represents the interactions of objects in a scene and their overall meaning. In this work, we propose a novel approach that integrates the video scene ontology with CNN (Convolutional Neural Network) for improved video tagging. Our method captures the content of a video by extracting the information from individual key frames. The key frames are then fed to a CNN based deep learning model to train its parameters. The trained parameters are used to generate the most frequent tags. Highly frequent tags are used to summarize the input video. The proposed technique is benchmarked on the most widely used dataset of video activities, namely, UCF-101. Our method managed to achieve an overall accuracy of 99.8% with an F1- score of 96.2%.