Unsupervised Image Generation and Manipulation Using Deep Convolutional Adversarial Networks

Abstract

In recent years, there has been an outburst in the field of Computer Vision due to the introduction of Convolutional Neural Networks. However, Convolutional Neural Networks have been sparsely used for unsupervised learning. The advancement of computational power and large datasets provide large opportunities to apply deep learning for image processing. This paper proposes a new architecture based on Deep Convolutional Generative Adversarial Network (DCGAN) for unsupervised image generation, its usage for image manipulation tasks such as denoising, super-resolution, and deconvolution. This proposed model demonstrates that the learned features can be used for image processing tasks—demonstrating their applications for general use as DCGAN learns from large datasets and adds high-level image details and features where traditional methods cannot be used. While the image results from the proposed network architecture and the original DCGAN architecture are similar in terms of performance, they are visually better when viewed by humans.