Speaker
Description
Colorizing grayscale images is a prominent challenge in Computer Vision, with numerous deep learning approaches being prevalent in the literature for this task. This paper introduces a multi-head fully convolutional network architecture, taking grayscale images as input and outputting the chromaticity information for colorization and segmentation masks for semantic segmentation. The model is inspired by the one proposed by Iizuka et al. in "Let there be Color!: Joint End-to-end Learning of Global and Local Image Priors for Automatic Image Colorization with Simultaneous Classification" (2016), but goes a step further by transitioning from iconic images to images with a number of different object categories. Preliminary empirical results, measured on both subjective and objective scales, demonstrate convincing colorization, without artifacts or color bleeding.
