Deep-learning-based super-resolution reconstruction of high-speed imaging in fluids

In many fluid experiments, we can only obtain low-spatial high-temporal resolution flow images and high-spatial low-temporal resolution flow images due to the limitation of high-speed imaging systems. To solve this problem, we proposed a degradation and super-resolution attention model (D-SRA) using unsupervised machine learning to super-resolution reconstruct high resolution (HR) time-resolved fluid images from coarse data. Unlike the prior research to increase the resolution of coarse data artificially generated by simple bilinear down-sampling, our model that consists of a degradation neural network and a super-resolution neural network aims to learn the mappings between experimental low-resolution data and corresponding HR data. What is more, channel and spatial attention modules are also adopted in D-SRA to facilitate the restoration of abundant and critical details of flow fields. The proposed model is validated by two high-speed schlieren experiments of under-expanded impinging supersonic jets. The comprehensive capability of D-SRA is statistically analyzed based on the synthetic unpaired schlieren images. The spatial-resolution of coarse images can be successfully augmented by 42 times and 82 times with most physical details recovered perfectly, which outperforms the existing method. The D-SRA also exhibits considerable generalization and robustness against unknown-degenerated schlieren images. Moreover, the practicability of the proposed method is also further explored on real unpaired jets schlieren images. It is convincingly demonstrated that the present study successfully surpasses the performance limitations of high-speed cameras and has significant applications in various fluid experiments to obtain flow images with high spatial and temporal resolution.