LiquidGAN: Integrating Liquid Neural Networks and Neural ODEs for High-Fidelity Thermal Image Synthesis Under Data Scarcity

The growing importance of thermal imaging in a variety of fields, including medical diagnostics, security, and environmental monitoring, emphasizes the critical need for efficient deep learning models capable of analysing thermal data. The training of these algorithms is much limited by the absence of suitably annotated thermal imaging datasets. We propose LiquidGAN, a novel Generative Adversarial Network (GAN) architecture which integrates the Neural Ordinary Differential Equations (ODEs) and Liquid Neural Networks (LNNs) to address the scarcity of annotated thermal imaging datasets. LiquidGAN employs a fixed-step Runge-Kutta (RK4) solver to refine the latent representations, enabling continuous image transformations through an encoder-ODE-decoder generator. The LNN-based discriminator adaptively models the subtle, and the dynamic temperature features, hence improving the ability to distinguish the real images from synthetic images. By integrating ODE-driven liquid dynamics across both the generator and discriminator, LiquidGAN effectively captures the dynamic temperature variations while stabilizing adversarial training under the limited data conditions. Moreover, Automatic Mixed Precision (AMP) and gradient scaling are used in our work to further improve the computational efficiency. Experimental results on multiple thermal imaging datasets, including those related to surveillance, healthcare, and precision agriculture, suggest that LiquidGAN offers a promising direction for generating high-fidelity synthetic thermal images under data-constrained conditions.