A Digital Superresolution Method With Minimal Sensitivity to Shift Estimation Error

Abstract

Superresolution (SR) methods become essential when an undersampled low-resolution (LR) image is unable to provide accurate target detection. The estimation of an HR image from a single LR image is ill-posed problem, and hence requires prior information. More the constraints, better is the reconstruction accuracy and this forms the basis of most of the contemporary state-of-the-art (SOTA) superresolution methods such as SRCNN and SRGAN which implement prior information as training set. Yet another approach to overcome the ill-posedness of the problem is to have multiple diverse LR images with the potential to reconstruct accurate HR image. Here we present the performance analysis of a generalized sampling theorem (GST) based multi-frame SR method. A simulation study using Gaussian targets is conducted, and a comparative performance analysis of the GST multi-frame SR method with the traditional multi-frame interpolation schemes and SOTA methods is presented using the percentage mean square error ( $\%$ MSE) and Structural Similarity Index Measure (SSIM). Our findings indicate that the GST SR method can outperform traditional interpolation and the SOTA methods.