Performance Analysis of Cycle GAN in Photo to Portrait Transfiguration Using Deep Learning Optimizers

Abstract

In the realm of computer vision, image transformations play a pivotal role across various domains such as healthcare, image enhancement, artist painting identification, genome sequencing, and more. While supervised learning demands a substantial volume of annotated images for training models, Cycle-GAN emerges as a potent solution for training models with fewer paired sources and target images in an unsupervised manner. This study introduces a novel system aimed at generating Monet-style paintings from realistic images, leveraging the Cycle-GAN methodology. Given the scarcity of Monet paintings, our system employs a combination of generator and discriminator neural networks to produce new Monet-style artworks. The model is trained using Cycle-GAN in conjunction with deep learning optimizers like RMSprop, ADAM, and SGD. The training dataset, Monet2Photo, comprises two distinct image categories: Monet paintings (300 samples) and natural photographs (7028 samples). The Monet-style images are utilized for training the model, while the raw photo images serve as the test set. Notably, the proposed model exhibits commendable performance, particularly when utilizing the SGD optimizer, as evidenced by favorable outcomes in terms of generator and discriminator losses.

Keywords

• Cycle-GAN
• generator
• discriminator
• object transfiguration
• unsupervised learning
• generative adversarial networks