Let there be Color!: Joint End-to-end Learning of Global and Local Image Priors for Automatic Image Colorization with Simultaneous Classification

摘要

We present a novel technique to automatically colorize grayscalernimages that combines both global priors and local image features.rnBased on Convolutional Neural Networks, our deep network featuresrna fusion layer that allows us to elegantly merge local informationrndependent on small image patches with global priors computedrnusing the entire image. The entire framework, including thernglobal and local priors as well as the colorization model, is trainedrnin an end-to-end fashion. Furthermore, our architecture can processrnimages of any resolution, unlike most existing approaches based onrnCNN. We leverage an existing large-scale scene classification databasernto train our model, exploiting the class labels of the dataset tornmore efficiently and discriminatively learn the global priors. Wernvalidate our approach with a user study and compare against thernstate of the art, where we show significant improvements. Furthermore,rnwe demonstrate our method extensively on many differentrntypes of images, including black-and-white photography from overrna hundred years ago, and show realistic colorizations.