Thermographic image super-resolution based on neural networks

The continuous development of thermographic technology has led to the overall improvement of instruments used in medicine, surveillance, or military systems. However, thermographic imaging cameras still have a high cost compared to other alternatives on the market, such as visible light cameras and a much lower spatial resolution. Super-resolution is a technique that improves the visual quality of an image through software processing. This work studies three neural networks architectures based on deep learning capable of performing super-resolution of RGB images at x2 and x4 scales: Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network (SRGAN), Enhanced Deep Residual Networks for Single Image Super-Resolution (EDSR), and Wide Activation for Efficient and Accurate Image Super-Resolution (WDSR). These architectures, in this work, have been trained as a super-resolution system using thermographic images as input data. The evaluation was carried out using thermographic images from different thermographic cameras. The performance assessment was carried out using the Peak Signal to Noise Ratio (PSNR) and the Structural Similarity Index Measure (SSIM). In addition, low-resolution images from a low-cost thermographic camera were used as input to the neural networks to study the feasibility of this method.