Project

DeepCAS

Novel GAN architecture using self-attention and conditional generation for high-fidelity satellite images.

Self-AttentionGANsLabel SmoothingDeep Learning

Overview

DeepCAS (Deep Conditional Attentional Smoothing) is a generative architecture that combines self-attention mechanisms, conditional generation, and label smoothing to produce high-fidelity satellite imagery. Published in Springer Nature 2025.

My Role

Co-author — contributed to the architecture design, attention mechanism implementation, and experimental evaluation.

Approach

Designed a conditional generator with self-attention layers for capturing long-range spatial dependencies in satellite imagery.
Applied label smoothing techniques to stabilize GAN training and reduce mode collapse.
Evaluated on the EuroSAT dataset using FID, IS, and classification accuracy metrics.
Demonstrated improved image fidelity and diversity compared to baseline GAN methods.

Outcome

Published in Springer Nature Singapore International Journal of Information Technology (2025). Outperformed existing methods on EuroSAT generation benchmarks.