A image-upscaling engine for 1080p to 4k using gradient-based interpolation

Abstract

ABSTRACT Image upscaling is an efficient solution to eliminate the gap between image/video sizes and display resolutions. Generally, image-upscaling algorithms use computation-intensive edge detection and filtering to improve their performance with inefficient hardware performances. In this study, an image-upscaling algorithm for 1080p to 4K using gradient-based interpolation technique is proposed. It simply adopts the horizontal and vertical gradients among source pixels to consider the texture content among them. This technique inherits the textures from source pixels and further extends them to target pixels using spatial-correlation. Experiment results demonstrate that this study has better average peak-to-noise-ratio (PSNR) and structural similarity (SSIM). The hardware architecture is realised using TSMC CMOS 0.18 μm technology. It comprises two core techniques: bubble-eliminating data scheduling (BEDS) and memory-efficient gradient generator (MEGG). BEDS can efficiently remove bubble cycles to improve hardware performances. MEGG can use compact memory capacity to produce gradient information. Its working frequency is 178 MHz with a power consumption of 9.43 mW. The maximum throughput is as high as 712 Mpixels/sec, which can sufficiently support 4K@60 fps. This study presents higher hardware efficiencies with better visual quality and object completeness in image upscaling for 1080p to 4K applications.