Skip to content

tandonmitul27/ColorizationUsingDNNs

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

73 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Colorization Of B/W Images Using DNNs

Overview

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.

Project Milestones

Week 1: MNIST Dataset and Linear Model

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.

Week 2: CNN Implementation and Advanced Techniques

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.

Week 3: Batch Normalization and Model Optimization

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.

Week 4: Research Paper Exploration and Model Iterations

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.

Model Iterations

Alpha Model

Trained a proof-of-concept model on a single image using autoencoders, demonstrating significant success in image retrieval.

Beta Model

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.

Final Model

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.

Conclusion

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!

About

WIDS'23

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published