Particle filter based multi-frame image super resolution

Abstract

In most multi-frame image super-resolution (SR) studies, the process that models the construction of the low-resolution (LR) images from high resolution (HR) one includes geometric transformation, blurring, downsampling, and adding noise. A simple linear approximation of this model is used in most super-resolution frameworks. The approximation error in this popular model appears as an additive noise that is generally a non-Gaussian noise. Most state-of-the-art estimators, such as the Kalman filter as a linear estimator used in super-resolution problems, suffer from this non-Gaussian noise caused by approximation error. We present a novel multi-frame super-resolution using the particle filter method that performs well in a non-Gaussian nonlinear dynamic system to produce the high-resolution output. One key innovation is that the approximation error in the super-resolution model can be handled well by using a particle filter. In this study, the optimal importance function is used to improve particle filter efficiency determined with the Taylor expansion of observation and measurement function. Experiments with simulated and real-image sequences yield that our proposed super-resolution method has good results in suppressing noise and reconstructing details in high-resolution images. © 2022, The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature.