Super-resolution reconstruction of schlieren images of supersonic free jets based on machine learning with bubble shadowgraphy data

Abstract

High resolution of supersonic flow fields is very relevant to advancement of supersonic and even hypersonic technology. Deep learning-based flow field super-resolution is a unique and efficient technique of reproducing high resolution of supersonic flow fields of free jet obtained from schlieren visualization. Enhanced super-resolution generative adversarial network (ESRGAN) and real enhanced super-resolution generative adversarial network (RealESRGAN) pretrained based on convolutional neural networks have been applied to enhance the spatial resolution of experimental supersonic flow field imaging by × 4 and × 8 without corresponding high-resolution ground truth images. The training dataset consists of high spatial resolution images acquired from bubble image shadowgraphy experimentation, while the test dataset was low spatial resolution schlieren images of supersonic free jet flow obtained from experiment at different nozzle pressure ratio (NPR) of 3–20. The pretrained networks were then applied to enhance low-resolution schlieren imaging of supersonic jet flow by increasing its spatial resolution via transfer learning with the network pretrained with bubble images and fine-tuned with patches of selected low-resolution schlieren images. The Natural Image Quality Evaluator (NIQE) score (no reference quality) was used to compute the perceptual quality of the super-resolved images without ground truth high-resolution images. The results showed impressive and improved flow field super-resolution with an average NIQE scores of 0.025–0.048 for ESRGAN × 4 and RealESRGAN × 4 and 0.026–0.053 for ESRGAN × 8 and RealESRGAN × 8, respectively.