Cross-Domain Image Processing Using Generative Models

Patricia Leonor Suárez Riofrío will defended her PhD thesis on March 24, 2026.

What is the thesis about?

Cross-domain image processing has emerged as a key research area in computer vision, driven by the need to overcome limitations posed by missing or inaccessible data from specialized sensors, which are often costly or unavailable. This field focuses on developing methods capable of transforming, complementing, or synthesizing image modalities from available visual data, emulating the human ability to infer non-visible information. Advances in deep neural networks, particularly Generative Adversarial Networks (GANs), have created new opportunities to address these tasks through data-driven learning techniques.

This thesis introduces a series of generative solutions designed to process images across different domains, including visible, thermal, near infrared (NIR), and depth representations, without relying on multispectral or specialized hardware. The work addresses monocular depth estimation from grayscale images to recover 3D structural information. It also explores cross-domain translation for thermal image synthesis from visible spectrum inputs and enhances this synthesis through multi-cue integration of depth and edge data.

Furthermore, it evaluates guided super-resolution techniques for thermal imagery using synthetic representations and proposes an advanced stacked conditional GAN architecture for NIR to RGB colorization. This work demonstrates that generative models provide a robust and effective framework for bridging modality gaps and enhancing representations that are otherwise difficult to acquire or access.

Keywords

Generative Models, Cross-Domain Image Processing, Cycle Generative Adversarial Network, Depth Map Estimation, Thermal Image Synthesis, Guided Super-Resolution, Multimodal Imaging, Deep Learning, Low-Cost Vision Systems.