Integrating regression model with Gaussian mixture model for image super-resolution

Abstract

The spatial resolution of the images captured by the optical components is very less and the image details are minimized due to problems, such as optical blurring, deviation in the lens and so on. Hence, the image resolution enhancing techniques have obtained more attention in recent years. This paper presents an image super-resolution (SR) method by integrating the Gaussian mixture model with the kernel regression model. At first, the low-resolution image is applied to the SR algorithm using GMM to obtain an HR image. Later, the high-resolution (HR) image obtained from the deep convolutional network is provided as the input to the kernel regression function to generate the enhanced high-resolution image. Finally, this paper analyses the performance of the proposed hybrid model for image super-resolution with the existing systems, such as Bicubic interpolation, SR using sparse representation of raw patches, Antipodally invariant metrics for fast regression-based super resolution, SR using joint GMM method using PSNR. Experimental results show, that the proposed model generates the enhanced HR image.