Multi-Modal Amharic Movie Genre Classification Using a Deep Learning Approach

The rapid growth of Ethiopia’s film industry, particularly in Amharic cinema, has created a demand for an intelligent model that can automatically classify movies by genre. Traditional manual methods are inefficient, inconsistent, and not scalable. Although various studies have explored movie genre classification using machine learning and deep learning on multimodal data, none have focused on Amharic films. This study proposes a novel deep learning-based multimodal model that combines video frames and audio data from Amharic movie trailers. It uses 10,152 video frame sequences and 13,169 audio chunks for training. Preprocessing includes Wiener filtering and spectral subtraction for audio enhancement, and CLAHE and Gaussian filtering for frame improvement. Frame features were extracted using a pre-trained I3D ConvNet, while audio features combined handcrafted descriptors (MFCC, ZCR, Chroma, Spectral Roll-off, and RMSE) with deep features from a BiLSTM network. Different neural network architectures, I3D, CNN, BiLSTM, and CNN-BiLSTM, were trained for action, comedy, drama, and romance genres using a multiclass classification approach and early fusion for feature integration. The models achieved accuracies of 89% (I3D), 77% (CNN), 82% (BiLSTM), and 92% (CNN-BiLSTM).In summary, the CNN-BiLSTM with early fusion outperformed unimodal models, demonstrating the effectiveness of combining visual and audio features for Amharic movie genre classification, thereby enhancing film organization, retrieval, and recommendation systems. The study demonstrates the effectiveness of multimodal fusion for low-resource languages and contributes to content organization, recommendation, and accessibility in Ethiopian cinema.