Using Clustered Frames to Classify Videos

This paper presents a semi-supervised learning technique to classify video clips. Usually, many tasks are done by categorizing video clips using deep learning techniques. However, based on the number of online videos today, it is necessary to use high computing power to accomplish this task. The authors propose methods that use Self-Organizing Map (SOM) to create a feature space representing clusters of video frames. The authors then classified them using simple voting, calculating entropy, neural networks, and Long-Short Term Memory (LSTM). The researchers also show finding frame numbers that are used to cluster video frames according to accuracy and training time. The results of this approach are presented based on testing 18 specific classes of real-world datasets from TV-programs containing 912 videos. The authors evaluated the techniques using five-fold cross-validation that our method archived 71.98% of average accuracy. Their computing time was then assessed, which achieved approximately 40 minutes of average computing time. Moreover, the researchers also compared the present proposal to other baseline models, including C3D and CNN-LSTM, and also used scene and action-recognition datasets, namely Hollywood2 to evaluate the technique. The authors archived 93.72% of average accuracy.