A Hybrid Approach to Movie Poster Similarity Detection: Combining Vision Transformers and Efficientnet for Enhanced Visual Recommendation Systems

This paper introduces a novel approach for analyzing movie poster similarity using an ensemble model that combines Vision Transformers (ViT) and EfficientNetB7. Traditional convolutional neural networks (CNNs) often struggle to capture global context and fine-grained details, which are essential for complex visual tasks like movie poster similarity detection. Our ensemble model leverages the global context modelling capabilities of ViT and the detailed feature extraction strengths of EfficientNetB7, resulting in a more comprehensive feature representation. The model is tested on a sub-dataset sampled from a dataset of 1,000 movie posters, each labelled with multiple genres. We construct similarity matrices for ViT, EfficientNetB7, and the ensemble model, followed by graph-based analysis to evaluate network properties such as density, clustering coefficient, and average shortest path length. The ensemble model outperforms the individual models, achieving a clustering coefficient of 0.8109, compared to 0.7807 for ViT and 0.7999 for EfficientNetB7, indicating a superior grouping of visually similar posters. The ensemble model demonstrates improved similarity detection by combining global and local feature extraction, making it suitable for realtime recommendation systems. Our findings provide new insights into the structural properties of similarity networks and demonstrate the effectiveness of hybrid models in visual recommendation tasks.