Image-Colorization

Automatic Image Colorization using a Convolutional Network (U-Net)

• Using the U-Net ConvNet Architecture for end-to-end image colorization.


• Takes as input a grayscale 32x32 image and returns a colorized 32x32 version


• The model has been trained on the CIFAR-10 32x32 images for 100 epochs.


• The model achieved an accuracy of 55.14% and a mean absolute error(MAE) of 0.0464 on the test set.

Model Achitecture

The model uses U-Net architecture which uses skip connections to preserve the lower level details and structute of an image, that are lost due to contracting bottle-neck.

The U-Net Architecture

Demo

A web interface has been implemented, where a user uploads a grayscale image as input and gets a colored image displayed as output

Sample Run

Requirements

• NumPy


• Tensorflow


• Keras


• SciPy


• Flask