DATA REFINEMENT USING GENERATED ADVERSARIAL NETWORK IN GEOPHYSICAL DATA

Abstract

This report presents a comprehensive explanation for an end-to-end Deep Neural Network for noise reduction,de-blurring and reconstruction of data based on im- age format. The learning is based on a Conditional Generated Adversarial Net- work(CGAN) and the content similarity that is exhibited, this loss is calculated in Wasserstein distance. This model achieves performance both in the structural similarity measure and generated to high real image accuracy. The quality of this model is also evaluated in a novel way on Geophysical problem { object detection on (de-)blurred and (de-)noised images and reconstructing whole large images us- ing GPU. The method is faster than algorithms that employ conventional algorithm and CPU computation. We introduce a method for generating synthetically blurred, noisy and pixelated images from sharp ones allowing real-life dataset augmentation, this in conjunction with geophysical data we are trying to use transfer learning and GAN for reducing the resources spend on training.