Multi-Frame Image Super-Resolution With Regularization

The limited number of high cost sensors in many cameras produce aliased images.In such cameras,the blur from the lens,color artifacts due to use of color filtering array,and noise from charged coupled devices(CCD)further degrade the quality of captured images.To mitigate these problems,the multi-frame super-resolution algorithms are designed to reconstruct high-resolution image by acquiring and fusing several low-resolution frames of the same scene beyond camera's resolution capacity.The early methods on super-resolution,although optimal for particular model of noise and data,generated poor results when implemented on real datasets.Meanwhile,single image super-resolution methods are introduced to reduce the color artifacts,but unable to eradicate such type of errors due to insufficient prior information.In this thesis,we propose an effective framework for the super-resolution problem to estimate the high-resolution image by using effective regularizations methods.Our framework addresses the main issues related to regularization and robustness in the image reconstruction.In this context,we propose the first novel edge preserving super-resolution model by using anisotropic diffusion process with₂ norm error minimization.In this method,we use statistical analysis to develop the new framework for diffusion process to perform regularization.The new framework of diffusion process leads to new diffusivity function that preserves sharper edges than previous formulations.Moreover,it improves automatic stopping of diffusion process which prevents the edges from over-smoothness.The proposed prior model is more effective in suppressing noises in the final reconstructed image than previous formulations.We propose second novel super-resolution model by using weighted sharpness total variation regularization.We employ robust₁ error norm minimization for the fast convergence rate in this model.In this super-resolution model,the regularization model develops the prior item by integrating the weighted sharpness of edges with total variation model to protect the edges with high sharpness.A complete implementation of super-resolution system is built in this thesis,and the experiments are performed on both real and synthesized datasets.The results obtained show the significant improvements over typical super-resolution methods in terms of average PSNR gains with 0.25 dB and 0.40 dB,and SSIM values with 0.025 and 0.056 in the first and second proposed works,respectively.