Dual-Branch Convolutional Neural Network for Image Comparison in Presentation Style Coherence

Image comparison is an important task that is part of the pipeline in many different computer vision applications. Maintaining style coherence across presentation slides is essential for professionalism and effective communication. Inconsistent design elements, such as varying fonts, colors, borders, and logo placements, can disrupt the visual flow and diminish the overall impact. This study introduces a novel approach to automate the validation of presentation slide coherence using a Dual-Branch Convolutional Neural Network. The model is trained to calculate a similarity score between image slides based on key design parameters, including font consistency, color schemes, border styles, and layout alignment. The proposed CNN architecture is specifically designed to compare two inputs representing slide images for binary classification. Unlike traditional Siamese networks that rely on identical branches and a distance metric for feature comparison, the proposed dual-branch architecture concatenates feature embeddings from two specialized branches and processes them through fully connected layers for final classification, allowing more targeted and nuanced feature extraction and coherence evaluation. The model was evaluated on a custom image dataset comprising 6000 images synthesized following specific design guidelines for style coherence of image features to ensure consistency and variety in the dataset while maintaining a balance for comparative tasks. The experimental results demonstrate significant improvements over the baseline Siamese network across all key metrics. Specifically, the proposed model achieved an accuracy of 0.85 compared to 0.81 for the baseline Siamese network, Jaccard similarity 0.76 vs 0.72, Kappa coefficient 0.69 vs 0.62, and ROC AUC 0.87 vs 0.81. Additionally, precision increased from 0.73 to 0.77 and the F1-score reached 0.87, reflecting a stronger balance between precision and recall. This work provides a significant contribution to automated design evaluation, offering a flexible and modular architecture that supports multi-view analysis and captures intricate visual patterns and discrepancies. By addressing key limitations of traditional approaches, the proposed model provides a robust tool to ensure style coherence in professional presentations, paving the way for more efficient and accurate design validation processes.