ICT4SS - ICT FOR SMART SOCIETIES
Image-to-image translation in adversarial latent space using deep neural networks
keywords DEEP LEARNING, VIDEO ANALYSIS
Reference persons ENRICO MAGLI
Thesis type RESEARCH
Description Convolutional neural networks (CNNs) have enjoyed great success in tasks such as image classification, object detection, etc thanks to weight sharing, invariances to geometric transformations and hierarchical
decompositions induced by the convolution operations. Recently, generative adversarial networks (GANs) have used CNNs to learn to generate new images by learning a mapping between a latent space where
random vectors are generated and a "real" image. Image-to-image translation is a class of problems where the goal is to learn the mapping between an input image and output image, translating an image from a given representation to another one. Converting a photograph into a painting of a famous artists, an aerial image into a map or an image of a horse into an image of a zebra are just a few examples of image-to-image translation problems. A new approach to solve these problems consists in using a Cycle GAN that enforces the mapping to be cycle consistent, in the sense that if we translate an image into another one and then translate it back, we should arrive back at the original image. The mapping obtained by a Cycle GAN is
defined in the space of pixels, not in the latent space. In this thesis the candidate will design a Cycle GANs that learns the mapping between input and output in the latent space, implementing the model in Tensorflow.
Required skills Candidate students should have a background in ICT or mathematics. Knowledge of neural networks (including TensorFlow environment and Python programming) is not a prerequisite, although it would help in the initial stage of the activity.
Deadline 20/11/2018 PROPONI LA TUA CANDIDATURA