This repository contains code and documentation for a comparative analysis of various image colorization techniques. The project explores both deep learning and non-deep learning methods, evaluating their performance on a dataset of grayscale images.
The repository is organized into several directories, each corresponding to a different colorization method:
SVR: Contains code and documentation for the Support Vector Regression method.Parzen_Window_Method: Code and documentation for the Parzen Window Method.CNN: Implementation details for the Convolutional Neural Network approach.cGAN: Information about the conditional Generative Adversarial Network (cGAN) method.
- Henry Lam
- Thomas Skøtt Gummesen
- Bartholomeus Diederik Rasmussen Pepping
- Magne Egede
The comparative analysis revealed various insights into the performance of different colorization techniques:
- SVR demonstrated suitability for smaller datasets but struggled with nuanced color transitions.
- The Parzen Window Method exhibited patchiness and inconsistent coloration, suggesting the need for refinement.
- CNN showed promising results, especially for nature images, but faced challenges with cityscapes.
- The cGAN method holds potential but requires further exploration and evaluation against other techniques.
Based on the findings, several recommendations for further research and development were proposed:
- Improving scalability and quality of SVR through alternative algorithms and smoothing techniques.
- Enhancing the Parzen Window Method with pixel-to-section level approaches and coherence optimization.
- Fine-tuning CNN models, exploring hyperparameter optimization, and incorporating diverse datasets.
- Further evaluation and comparison of cGAN against other methods, considering training data quantity and quality.
For detailed insights and analysis, please refer to the full written report available in my blog.
For questions, feedback, or support, please contact me at support@henrylam.blog.