Automatic Colorization with Colourfulness metric using Deep Neural Networks

Abstract

Colorization is the automated procedure that applies colors on a monochrome image or sequence of images. A model in this colorization process will segment an image into regions or blocks and trace out those blocks across the total picture. Practically the best colorization procedures need human intervention and also cost-effective. The enhanced process must be completely automated colorization methods still raises different complications in the research aspect. In this paper, we have described two approaches that are helpful during colorization. Initially, we evaluated image color useful in compression of multiple images, camera sensor modules, and to obtain the aesthetic quality of images. By using the montage tool, we separated the colorfulness of the pictures from extremely colorful to least colorful. Besides, we presented the scales and measures for finding the image colorfulness quality.  Next, unlike the traditional methods, we used a deep learning process using a pre-trained CNN for colorizing the monochrome or black and white images. This model is a pre-trained Caffe model trained on billions of images and examined with the sample images. The transformation of RGB images to L*A*B color space yields better results in contrast with state-of-art methods.