Single Frame Super-resolution Reconstruction Method Based on Small Convolutional Recurrent Neural Network

Abstract

The invention discloses a small convolutional recurrent neural network-based single-frame super-resolution reconstruction method. The method comprises the following steps of: establishing a recurrentconvolutional neural network for linearly activating functions, wherein an input of the recurrent convolutional neural network is a small neighborhood of a certain pixel point of an image and an output of the neighborhood is a corresponding temporary output result at the pixel position; and inputting a low-resolution image by utilizing the established neighborhood, and carrying out iteration for certain times so as to output a final super-resolution imaging result. Compared with existing statistical learning-based image super-resolution method, the method nearly can ignore the model complexityand calculated amount, and has more specific physical meaning for internal parameters. External data is imported to assist corresponding model parameters to carry out learning, so that the method iscapable of obtaining a relatively good super-resolution reconstruction effect.

Description

technical field [0001] The invention relates to a convolutional recurrent neural network and computer image processing calculations, in particular to a single-frame super-resolution reconstruction method based on a small convolutional recurrent neural network. Background technique [0002] With the development of information technology and the promotion of smart devices, users have higher and higher requirements for image quality. However, due to the aberration of the optical imaging system, the bandwidth limitation of the image acquisition device itself, and the bandwidth limitation in the transmission process, the image quality of the image cannot be improved without limit. Therefore, how to adopt a suitable super-resolution algorithm to improve image quality and overcome the above-mentioned problems has become a very popular research topic. Hardware-based super-resolution usually has higher precision and better performance, but its application is often limited and the co...

AI Technical Summary

Problems solved by technology

Hardware implementation of super-resolution usually has higher precision and better performance, but its application is often limited and the cost is high, and it cannot be applied to other imaging environments

[0003] There are two existing software implementations of super-resolution algorithms. The first one is based on multi-frame super-resolution reconstruction. Starting from the image degradation model, multiple low-resolution images are registered at the sub-pixel level and reconstructed using the registration information. However, this method requires that the exposure and other factors between multiple low-resolution images are exactly the same, and the registration accuracy is extremely high, which greatly limits the application of this method, and the single-frame super-resolution Algorithms often have complex models, rely on a large number of external training samples, and have extremely high computational complexity.

Images