Seminar: Deep Neural Network for Image Stylisation

Zili Yi
Ph.D. Oral Comprehensive
Supervisory Committee: Dr. Minglun Gong, Dr. Wolfgang Banzhaf
and Dr. Yuanzhu Chen

Deep Neural Network for Image Stylisation

Department of Computer Science
Thursday, April 6, 9:30a.m., Room EN 2022


Abstract

Image stylisation aims to automatically generate stylised images by manipulating photographs. Loosely, image stylisation includes any efforts that try to generate patternized images through manipulating or rendering input digital photographs, coveringa broad range ofsubjects varying from artistic stylisation to naturalistic rendering, to tone harmonisation, colourisation, abstraction, image filtering, super-resolution and even stylised handwriting generation. A narrower definition of stylisation onlyrefers to artistic stylisation, also known as non-photorealistic rendering. Inmythesis, I utilise the following generalised definition of image stylisation.

The generalised problem of image stylisation can be formatted as, given an input image or image sequence, and a targeted visual style either described in natural language or illustrated with one or multiple examples, design a system that could automatically output an image or image sequence of the targeted visual style and containing the contents of the input, while assuring robustness to variation of contents, image quality, lightening condition and source of the input. Based on this statement, a wide variety of problems such as non-photorealistic rendering, tone harmonisation, black-white image colourisation, image abstraction, image filtering, image denoising, image super- resolution, and augmented reality, can be formatted as image stylisation.

Traditional methods for image stylisation usually involve a sequence of manipulations and can be unified as a workflow of processing unit, where each processing unit is deliberately designed or consideratelyfine-tuned for specific applications. However, Deep Neural Network (DNN), especially the recently-emerging Fully Convolutional Networks (CNN) and Generative Adversarial Network (GAN), which take an image as the input and output another image, enables powerful end-to-end architecture to be applied to image stylisation. In addition, image stylisation can also benefit from the advanced cognitive power and exceptional modelling capability of DNN.