Deep Learning for Enhanced Audio-Based Movie Scene Classification via Spectrogram Analysis

This study investigates the feasibility of employing deep learning models for the classification of movie scene audio into distinct categories using spectrogram analysis. To facilitate this, a dataset comprising 766 30-s audio clips was developed, categorized into six distinct classes: Action, Comedy, Horror, House, Racing, and Sport. These audio files were transformed into Mel-spectrograms to serve as visual input for various pre-trained deep learning architectures. The study leveraged transfer learning with eight established MATLAB pre-trained models (DarkNet-19, DarkNet-53, Inception-v3, MobileNet-v2, ResNet-18, VGG-19, VGGish, and YAMNet) to assess their comparative performance. Experimental validation on unseen data revealed that the VGGish model demonstrated superior classification capability, achieving the highest accuracy of 83.00%. This level of accuracy from VGGish, alongside notable performances from models like MobileNet-v2 (80.00%) and DarkNet-19 (80.30%), underscores the significant potential of using pre-trained Convolutional Neural Networks (CNNs) for this task. The findings suggest that such models, particularly VGGish which is pre-trained on audio data, can effectively learn discriminative features from spectrograms, offering an efficient and promising pathway for automated audio-based movie scene classification. This approach contributes to enhancing content analysis and retrieval in multimedia applications.