Reconstruction and regularization multi frame super resolution on vegetation index NIR image

Vegetation index is measured using remote sensing with VNIR image acquired by satellites, one of them is PROBA-V. It gathers pair of low resolution (LR) images and high resolution (HR) images. The LR images is acquired faster but contains aliasing. Hence it can be processed into high resolution image using multi frame super resolution. But, to have an ideal LR image as a comparison, new synthetic LR image dataset is generated using only translation, gaussian PSF, and gaussian noise. Two type of approaches are used, reconstruction and regularization. Results from both methods are post-processed using median filter to remove noise due to error in super resolution process and poorly chosen hyperparameter. Then, the result is evaluated using PSNR and SSIM by compared to ground truth from dataset HR images. Also, simple bicubic interpolation is used to measure any information improvement by performing super resolution. For both LR images from dataset and synthesis, highest PSNR and SSIM are provided by regularization method due to its multiple iteration for predicting high resolution image, meanwhile reconstruction method only uses single iteration.