Real-time super-resolution method and system based on FPGA

Abstract

The invention discloses a real-time super-resolution method and system based on an FPGA, and relates to the technical field of image processing. The method comprises the steps of splitting each frameof image in a medium with a relatively low resolution into sub-images for performing super-resolution processing; allocating a processing module according to an eigenvalue of the current frame of thesub-image; if the eigenvalue is relatively high, performing calculation by a neural network module; and if the eigenvalue is relatively low, performing calculation by an interpolation module. An output buffer outputs the high-resolution sub-images obtained by calculation and recombines the sub-images into a high-resolution image, so that a user can see the real-time ultra-high-definition image.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a field-programmable gate array (FPGA)-based real-time super-resolution method and system. Background technique [0002] Ultra High Definition (UHD) technology is dramatically changing the media industry, with high-resolution displays now dominating the market. However, the supply of ultra-high-definition media resources is limited, and access is difficult due to insufficient network bandwidth, and the actual user experience is still mediocre. Therefore, it is of great significance to efficiently upscale 1920×1080 traditional high-definition (2K FHD) video to 4K UHD resolution (3940×2160). The technique of estimating fine-grained resolution images / videos from low-resolution inputs is often referred to as super-resolution, and this fundamental problem is very attractive in the fields of image processing and computer vision. [0003] Field-programmable gate array (FPGA)-...

AI Technical Summary

Problems solved by technology

Model-based methods aim to recover high-resolution scenes based on observation models and priors (regularization). Most of the existing techniques assume known blur kernels and noise levels, but in practice the blur kernels and noise levels can be arbitrary. , assuming that it is known to cause difficulties in practical applications, the effect is not good

Neural network-based super-resolution often has a high computational cost, and often faces the problem of insufficient output frame rate when processing video in real time

Images