بهبود توأم بازيابي جزئيات و كاهش اثر نويز در ابرتفكيك پذيري چندقابي

Image Super-resolution , Image reconstruction , Particle filter , Frame selec‎tion , Image channels

Joint Reconstruction an‎d noise reduction in multi frame Super Resolution

مهندسي برق

Multi-frame Image Super-Resolution (MFSR) is one of the advanced an‎d efficient approaches for reconstructing high-resolution images from a set of low-resolution frames, widely used in critical applications such as satellite imaging, video surveillance, an‎d medical imaging. However, this approach faces fundamental challenges, including noise caused by model approximation errors, geometric variations between frames, an‎d high computational deman‎ds. This dissertation aims to address these challenges by presenting an innovative framework for joint image reconstruction an‎d noise reduction in the context of multi-frame super-resolution. This research addresses the SR problem in two parts by reducing the effect of noise: in the first part, by considering the modeling error in the form of non-Gaussian noise an‎d using particle filters for SR. In the second part, finding the best finite frames based on increasing the signal-to-noise ratio an‎d reducing the noise caused by modeling the motion function between frames. Thus, this research is presented in two parts. In the first part of this study, a novel particle-filter-based method for high-resolution image reconstruction is developed. The conventional modeling of the low-resolution image formation process typically involves geometric transformations, blurring, downsampling, an‎d noise addition, which are often approximated linearly. Such approximations introduce model errors that manifest as non-Gaussian noise. The use of the Kalman filter, as a linear estimator, proves inefficient under these conditions. In this research, by employing a particle filter an‎d selec‎ting an optimal importance function based on a Taylor expansion, a method is proposed that can effectively operate in nonlinear dynamic systems with non-Gaussian noise. Experimental results on both simulated an‎d real datasets demonstrate that the proposed method is capable of preserving finer details in the reconstructed image an‎d in the simulated sequence, it effectively reduces non-Gaussian noise by close to 1.4 dB. In the second part, to reduce computational load an‎d memory consumption in multi-frame super-resolution frameworks, a new frame selec‎tion algorithm is proposed. This algorithm focuses on selec‎ting an optimal subset of non-redundant frames, aiming to identify the minimum number of frames that contain the most useful information for image reconstruction. The proposed method, through multi-channel sampling, reference frame selec‎tion, an‎d signal-to-noise ratio optimization, implements two search algorithms. In the first step, cases with the maximum number of non-empty channels are identified, an‎d in the second step, for each channel, the frame with the minimum model transfer noise is selec‎ted. A set of frames with the lowest overall noise is then extracted an‎d used as input for the reconstruction process. The experimental eva‎luation of the SR methods presented in this thesis shows that the proposed method is very fast an‎d lightweight, an‎d in simulated experiments an‎d real frame sequences, using SR methods along with the proposed frame selec‎tion algorithm reduces the execution time while maintaining the quality of the output image compared to all frames, an‎d in simulated sequence experiments, there is an increase in quality of nearly 2 dB compared to selec‎ting the first few frames.

7

162674