Neural network classification of videos based on a small number of frames

The article proposes a method for neural network classification of short videos. The classification problem is considered from the point of view of reducing the number of operations required to categorize videos. The proposed solution consists of using a small number of frames (no more than 10) to perform classification using the lightest neural network architecture of the ResNet family of models. As part of the work, a proprietary training dataset was created, consisting of three classes: "animals", "cars" and "people". As a result, a classification accuracy of 79% was obtained, a database of classified videos was formed, and an application with GUI elements was developed for interacting with the classifier and viewing the results.