Welcome to the Colorization Project! This project focuses on the development of a colorization model using a variety of powerful libraries and techniques. We leverage the capabilities of Numpy, Pandas, PyTorch, and Matplotlib to enhance our understanding of colorization processes and achieve remarkable outcomes.
Loaded, preprocessed, and augmented MNIST dataset using Numpy and PyTorch. Trained a model with Linear layers achieving an accuracy of 89% in recognition of handwritten digits.
Implemented a CNN using PyTorch for improved feature extraction and model performance. Explored custom loss functions, focusing on perceptual losses for colorization. Introduced transfer learning for future model development.
Implemented batch normalization techniques to stabilize training and address gradient-related issues. Achieved an impressive accuracy of over 96% on the MNIST dataset using advanced techniques.
Studied a Medium article by Emil Wallner on Colorize B&W Photos with a 100-line Neural Network. Understood Convolutional Autoencoder architecture, a crucial concept for subsequent implementation phases. Explored a research paper on the colorization of black and white images for valuable insights.
Trained a proof-of-concept model on a single image using autoencoders, demonstrating significant success in image retrieval.
Expanded the scope by training the model on multiple images and increasing the number of epochs. Structured the use of encoder and decoder components, contributing to improved colorization outcomes.
Leveraged transfer learning by integrating the Inception ResNet v2, a powerful classifier trained on 1.2M images. Transferred learning from the classifier to the coloring network, enhancing the model's understanding of object representations.
The success of the alpha model, scalability in the beta model, and the integration of transfer learning in the final model collectively demonstrate a progressive and well-rounded approach to achieving the project objectives. This README serves as a guide to understanding the journey, techniques, and accomplishments of the Colorization Project. Feel free to explore the code and contribute to further advancements in colorization technology!