Overview / Usage

Generative Adversarial Network (GAN) can produce images which are strikingly similar to the data provided with great visual quality. This paper reports on an attempt to generate realistic underwater images using a GAN. Our main purpose is to evaluate the proposed model to generate underwater images which resemble real world samples.

Methodology / Approach

The initial step was data collection. Web image hosting service Flickr has been our primary source for images. We also contacted few professional photographers and used some of their images. Through all these sources, we came up with an image dataset containing close to 5000 images as our training dataset.

We use a Deep Convolutional GAN (DCGAN) implementation along with the Wasserstein GAN algorithm for the model.