Pseudo-Cyclic Network for Unsupervised Colorization with Handcrafted Translation and Output Spatial Pyramids
Résumé
We present a novel pseudo-cyclic adversarial learning approach for unsupervised colorization of grayscale images. We investigate the use of a non-trainable, lightweight and well-defined Handcrafted Translation to enforce the generation of realistic images and replace one of the two deep convolutional generative adversarial neural networks classically used in cyclic models. Additionally, we propose to use Output Spatial Pyramids to jointly constrain the deep latent spaces of an encoder-decoder generator to preserve spatial structures and improve the quality of the generated images. We demonstrate the interest of our approach compared with the state of the art on standard datasets (paintings, landscapes, aerial, thumbnails) that we modified for the purpose of colorization. We evaluate colorization quality of the generated images along the training with deterministic and reproducible criteria. In complement, we demonstrate the ability of our method to generate representations that are prone to make a classification network generalize well to slightly different color spaces. We believe our approach has potential applications in arts and cultural heritage to produce alternative representations without requiring paired data.