Combining Computer Vision and Word Processing to Classify Film Genres

The paper investigates various deep learning approaches for classifying movies by genre. To prepare the initial dataset, a system for parsing textual information and images from the Internet Movie Database, IMDB (https://www.imdb.com/) was implemented. Pandas (Python library) methods were applied to generate the dataset. Particular attention was paid to the preparation of target information and the choice of the learning loss function due to the fact that the problem was reduced to a multi-label classification. A number of transfer learning algorithms have been implemented for convolutional neural network models and transformer architectures when processing movie posters, as well as simple machine learning models and deep learning transformer-type models for text classification. The paper proposes a method for complexing extracted features from heterogeneous data in another neural network to improve classification accuracy. It was possible to achieve an increase in the average cosine distance for the test sample by 3% higher than when using standard models and about 7% in sense of intersection over union metric.