Auto-Colorization of Historical Black and White Images
Introduction
The project entails both a re-implementation and innovative expansion of the paper titled “Colorful Image Colorization” authored by Richard Zhang, Phillip Isola, and Alexei A. Efros to develop an automated colorization model for historical black and white images, utilizing Convolutional Neural Networks. In contrast to traditional image processing and computer vision techniques employed in prior works, the paper presents a CNN-based approach equipped with a specialized loss function tailored specifically to the colorization problem, generating more convincing and perceptually realistic colorized images. Our aim is to replicate and build upon the achievements of the original research by incorporating Gaussian Mixture Models and a unique CNN architecture, with a primary focus on addressing and enhancing the challenging multi-modal problem. Developers involved in this project will comprehend and analyze various academic papers, design and implement deep learning architectures, and collaborate with fellow teammates to achieve improvements on the quality of generated images.
We are actively seeking highly self-motivated developers with rich experience in Python, a strong foundation in Linear Algebra and Statistics, and adequate knowledge in Deep Learning. Proficiency in key relevant frameworks such as PyTorch, NumPy, Keras, and TensorFlow is essential. Additionally, expertise in Image Processing, Computer Vision, and Pattern Recognition will be highly advantageous. Please email ruihang.zhang@mail.utoronto.ca and xiaoo.zhang@mail.utoronto.ca if this opportunity aligns with your interests and you would like to explore further project details!
