Self-adaptive super-resolution method based on meta transfer learning

Abstract

The invention provides a self-adaptive super-resolution method based on meta transfer learning, which comprises the following steps: pre-training a self-adaptive super-resolution model based on meta transfer learning through an external image data set, so that the model can learn prior information of image reconstruction; performing random parameter down-sampling on images in the external image data set through a random Gaussian sampling method, so that training data in a meta transfer learning process contains multi-task information; performing down-sampling on a to-be-reconstructed target low-resolution image to obtain a low-resolution sub-image, and training the model by taking the low-resolution image and the low-resolution sub-image as training data; performing adversarial training by means of a twin neural network and the model, and adjusting model parameters by comparing the difference between a low-resolution image and a low-resolution sub-image to complete model training; and applying the model to a target image to reconstruct and generate a super-resolution image. According to the method, the image reconstruction quality can be improved, and the generalization ability of the super-resolution model is enhanced.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to an adaptive super-resolution method based on meta-transfer learning. Background technique [0002] The concept of super-resolution was first proposed in the 1960s. Super-resolution initially only meant the restoration of a single image, and various image restoration methods appeared later, but this technology has not been widely used in practice. At the end of the 20th century, with the development of computer technology, signal processing theory and optimization theory, a series of super-resolution methods have been proposed, mainly including interpolation-based, reconstruction-based and learning-based methods. Today, the vigorous development of deep learning technology has brought the performance of image super-resolution to a higher level. The current research on super-resolution includes image super-resolution based on interpolation, image super-resolution b...

AI Technical Summary

Problems solved by technology

[0002] The concept of super-resolution was first proposed in the 1960s. Super-resolution initially only meant the restoration of a single image, and then various image restoration methods appeared, but this technology has not been widely used in practice.

[0007] (1) Most of the focus is on the design of the model structure and the data set, and many methods ignore the computational burden brought by the deep neural network. Performance on actual super-resolution tasks is often not as good as on benchmark datasets

[0008] (2) Although many deep learning models have achieved better reconstruction results, deeper networks also bring about problems such as overfitting and slow convergence speed, which are common problems of deep neural networks. Most of the high super-resolution results depend on the repeated parameter adjustment of the network, and the final model is difficult to reproduce in actual application scenarios

Images