Infrared Image Transformation via Spatial Propagation Network

In recent years, there has been an increasing demand for intelligent infrared recognition methods. As current high-precision intelligent recognition algorithms like deep networks largely rely on massive amounts of training data, the lack of infrared databases has become a major limitation for technological development, resulting in an urgent demand for intelligent infrared image simulation technology. Different from most infrared image simulation techniques which expand infrared data amount under conditions of thermal balance, this paper proposes a novel way to simulate infrared images, i.e. generating infrared images for objects in scenes under an unsteady heat conduction process along the time axis. To be specific, this paper incorporates a spatial propagation network structure to predict the equivalent thermal conductivity coefficients for the input infrared image captured at a certain time point, and then infers the infrared images at the next time points by simulating the physical heat conduction process based on the predicted conductivity coefficients. We carry out extensive experiments and analysis on the datasets composed of factual infrared photos and PDE-simulated images, demonstrating that the proposed infrared image generation method can realize the transformation simulation and dataset expansion of infrared images with high speed and high quality.