Movie Scene Genre Classification & Video Recommendation System using Deep Learning

Abstract

Object and scene classification methods have often struggled to match human performance due to background complexity, context understanding and complex foreground elements that often appear blurry or moving. These are particularly evident in movie analysis with dynamic backgrounds and frames containing complex features. With the recent advancement and development of deep learning and convolutional neural networks (CNN), numerous possibilities have open for better scene categorization and higher-level tasks such as genre classification. This study aims to build up on the ---- and proposes a transfer learning based CNN model that utilizes inception V3 model. the study utilized 80 videos from YouTube that contained movie scenes over four genre types; Action, Horror, Sci-Fi and Romance. The extracted scenes from each video were analyzed and Euclidean Distance method was applied for measuring dissimilarity and to remove any redundant scenes. The model was cross tested on unseen data and evaluated using Precision, Recall, Accuracy and F1 score. The base line model achieved a 66% accuracy which was further improved by fine-tuning the mode to 72%. Furthermore, the model achieved a 67% Precision, 64% Recall and a 65% F1 score. The result demonstrated that the proposed transfer learned model showed accuracy surpassing human classification abilities. The findings of the research can further aid in the applications related to content tagging, recommendation systems, film analytics and copyright monitoring, by providing a faster and scalable genre classification system.

Keywords: Deep Neural Network, Transfer Learning, Movie Recommendation, Scenes Detection