A multi-frame super resolution image reconstruction using regularization framework

Abstract

The global world experiences an enormous advancement in software and hardware technologies within the past decade. One of the primary measurements of image quality is image resolution. High-resolution is generally required and preferred for producing more detailed information inside the digital images; therefore, this leads to improve the pictorial information for human analysis and interpretation and to enhance the automatic machine perception. Unfortunately, the best use of image sensors and optical technologies is usually a high-priced method and is also constrained to increase the image resolution. Therefore, the effective use of image processing techniques for acquiring a high-resolution image generated from low-resolution images is an inexpensive and a powerful solution, which is called multi-frame super-resolution image reconstruction. However, the real imaging systems may introduce some degradation or artifacts in the digital images. These distortions in the images are caused by a variety of factors such as blurring, aliasing, and noise, which may affect the resolution of imaging systems and produce low-resolution images. Numerous strategies like frequency and spatial domain approaches have been proposed in the literature. Spatial domain approaches are classified as one of the most popular approaches and split into interpolation-based approaches and regularization-based approaches. Nevertheless, these techniques still suffer from artifacts. Regularization-based approaches are a challenging in image super-resolution in the last decade. This research intends to enhance the efficiency of multi-frame super-resolution image reconstruction in order to optimize both analysis and human interpretation processes by improving the pictorial information and enhancing the automatic machine perception. As a result, this research proposes new approaches for the image reconstruction of multi-frame super resolution, so that they are created through the use of the regularization framework. On one hand, an efficient proposed approach is derived from the hybrid of reconstruction models in the image reconstruction stage based on employing adaptive nonn and Lp norm in the data-fidelity term, beside adopting bilateral edge preserving and bilateral total variation prior models in the regularization term respectively. On the other hand, an efficient initialization approach is based on estimating the initial high resolution image through the pre-processing stage on the reference low resolution image. The proposed initialization approaches use linear and nonlinear filters including median, mean, Lucy-Richardson, and Wiener filters at the reference low resolution image. The proposed approaches are used to approximate a high-resolution image generated from a sequence of corresponding images with low-resolution to protect significant features of an image such as sharp image edges and texture information while preventing artifacts. In addition, these proposed approaches generate a high-quality image that is used in realistic applications with edges preservation and noise suppression. The experimental results with synthetic data indicate that the proposed approaches have enhanced efficiency visually and quantitatively compared to other existing approaches.