Work

Hybrid GAN and Fourier Transformation for SAR Ocean Pattern Image Augmentation

Year: 2024

Type: conference-paper

Abstract: Synthetic Aperture Radar (SAR) images provide valuable information for ocean observation and analysis, aiding the understanding of oceanic processes and their role in climate change. This study presents a Generative Adversarial Networks (GANs)-based approach for generating realistic and diverse ocean-pattern SAR images. The proposed methodology combines a style-based generator network with an adversarial discrimina-tor network to learn and reproduce complex patterns present in SAR images. To avoid discriminator overfitting due to insufficient training data, an adaptive discriminator augmentation (ADA) mechanism has been exploited. However, the generated images lacked physical properties in the frequency domain. To address this issue, we modified the loss functions by adding the mean square error of the generated and real images in the frequency domain, improving the spectrum of the generated images.

Source: IEEE International Workshop on Metrology for the Sea; Learning to Measure Sea Health Parameters (MetroSea)

DOI: 10.1109/metrosea62823.2024.10765638

Authors: Mobina Keymasi, Omid Ghozatlou, Emmanuel Weridongha Adueze, Mihai Datcu

Institutions: Universitatea Națională de Știință și Tehnologie Politehnica București, Constanta Maritime University

Language: English

Topic: Generative Adversarial Networks and Image Synthesis

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

Funder: European Commission

Award: European Training Network (ETN) on Multimodal Environmental Exploration Systems – Novel Technologies