Deep learning for style and domain transfer

• Main Author: Ni, Anqi
• Other Authors: Wen Bihan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/158046
• Institution: Nanyang Technological University

The diversity of painting styles provides rich visual information for constructing artistic images. In this project, two image style transfer algorithms based on deep learning are proposed and tried. One is CNN-based algorithm, which uses pre-trained convolutional neural network (CNN) to extract the features of each layer of the network, separates and reorganizes the content image and style image, and constructs a new loss function to obtain a new artistic style image. Another algorithm is based on generative adversarial network (GAN), which can directly translate an image between the source and target domains. Using cycleGAN as baseline, new artistic style pictures are obtained by new proposed generators. The experimental results show that the new images generated by the two models have their own advantages and disadvantages, but both can achieve good style transfer results. The deep learning-based image style transfer algorithm and models proposed in this project constructs richer visual information and also provides a reference for new artistic creations.