Development of Efficient Image Upscaling Techniques

Abstract

Image upscaling is a popular topic in recent years, and it is used to create a high resolution (aHR) image from low resolution (LR) image data. An efficient image upscaling approach must preserve the original LR image’s edge information, texture, geometrical regularities, and smoothness while producing its HR counterpart. The most typical use of image upscaling is to improve the visual effect of a digital image after resizing it for displaying and printing. Image upscaling uses a variety of polynomial interpolation algorithms due to their low computational complexity and applicability for a wide range of real-time applications. A polynomial interpolation approach uses the weighted average or convolution of surrounding pixels to obtain the interpolated value at a specific place. This might cause blurring effects in upscaled images due to high frequency deterioration. This issue can be addressed by utilizing edge-directed algorithms that maintain high frequency information in an upscaled image for improved visual quality. Although edge-directed interpolation strategies are effective at preserving fine details and edge information in an image during upscaling, they are computationally more complex than polynomial interpolation schemes due to the usage of adaptive and local-based techniques. Most transform-domain interpolation algorithms in the literature produce blurring effects in upscaled images, particularly at edges and high-variance regions. Learning-based picture interpolation algorithms can produce high-quality results with fine features, but they often require significant computing resources and training data. This dissertation suggests pre-processing approaches to reduce blurring effects in upscaled images, incorporating an improved discrete cosine transform that recovers lost information due to upscaling using a bilateral filter. The weighted missing details are then integrated with the LR image before interpolation, resulting in less blurring in the high-variance region. However, in the following method, a higher order Laplacian filter is used to sharpen the edge presence in each direction of the degraded image in order to predict small details before combining them into an LR image. This strategy reduces blurring caused by interpolation. However, with iterative optimization sharpening, missing details are sharpened repeatedly using an optimal filter before interpolation, resulting in a better recovered HR image with more detailed information. However, an effective method of image upscaling is suggested here that combines iterative-back projection to reduce blurring effects with a convolutional neural network to retrieve both shallow and deep data separately. Some post-processing solutions have also been proposed, including an improved transform-domain approach that employs discrete sine transform upscaling to improve the quality of the HR image via difference image after upscaling. Another cubic B-spline spatial-domain approach involves sharpening degraded high frequency data and combining it with an upscaled image to get the restored HR image. Then comes a new technique: adaptive edge sharpening-optimized directional anisotropic diffusion, in which the smooth and edge areas are treated individually after upscaling to eliminate blurring effects and improve fine details. Several hybrid techniques have been developed to reduce the blurring effect in interpolated images. Hybrid approaches are developed by merging pre- and post-processing algorithms. In optimal local adaptive edge preserving spline, the high frequency details of an LR image are increased to compensate for blurring in the equivalent upscaled image. Furthermore, edge expansion is used to anticipate high frequency features with local statistics, preserving the edge contents in the upscaled image. However, the edge-error (EE) method uses the LR image’s edge to guide interpolation, whereas the edge-residual (ER) approach uses both the LR edge and lost information, followed by sharpening with a higher order filter. The restored HR image is generated by combining the sharpened image and the interpolated image.