Infrared Image Super-Resolution via Generative Adversarial Network with Gradient Penalty Loss

Infrared thermal imaging technology has been gradually developed and widely applied in measurement and non-destructive testing. However, low-contrast blurred details and expensive acquisition equipment remain as barriers to its further practical applications and widespread adoption. In this paper, a novel framework comprising deep learning techniques is proposed to offer a relatively competitive and compatible solution of infrared image super-resolution. Firstly, radiance information from low-resolution imagery is detected and automatically translated to high-resolution through a Generative Adversarial Network (GAN) with Wasserstein distance. Secondly, a gradient penalty loss function is utilized for the discriminator to guide the generator to achieve reasonable and acceptable convergence. Through evaluation of three widely utilized infrared datasets, the proposed method demonstrates superior performance against the state-of-art method with more accurate Peak Signal-To-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) respectively. The outcome of this study has implications for a real-application of deep learning based infrared non-destructive testing and measurement scenarios.