Image Colorization Using Convolutional Neural Networks

Abstract

Today, colorization is done by hand in Photoshop. A picture can take up to one month to colorize. It requires extensive research. A face alone needs up to 20 layers of pink, green and blue shades to get it just right. Coloring grey scale images manually is a slow and hectic process. Using machine learning techniques this can be done very fast. Convolutional neural networks have emerged as a standard in image classification problems. They achieve higher accuracy rates over all the techniques in detecting patterns in images. As this problem mostly deals with identifying the pattern in the image and colorizing it accordingly convolutional neural networks serves the best. In this project we combines a deep Convolutional Neural Network trained from scratch with high-level features extracted from InceptionResNet-v2 pre-trained model. The purpose of our model is to estimate a* and b* components from the luminous component, L of the input image. The a* and b* components of the reconstructed image are combined with luminous component to give the estimated color image. The convolutional layers are a set of small learnable filter that help us identify patterns in an image. The layers close to input layer looks for simple patterns such as edges and outlines, and layers close to output layer looks for cowmplex patterns.