Super-resolution amplification method, system and device for image with any scale

Abstract

The invention provides a super-resolution amplification method, system and device for an image with any scale and a storage medium, and the method comprises the steps: collecting an original image data set and down-sampling images with different multiples, and carrying out the paired storage of the image data set and the down-sampling images with different multiples, so as to serve as a training data set; inputting the down-sampling image and the super-resolution multiple into a training model to obtain a feature map of the down-sampling image, and performing mask filling on the feature map according to the super-resolution multiple to form a new feature map; mapping the new feature map to a super-resolution image, calculating the loss between the super-resolution image and the original image, and updating the training model according to the loss until the training is completed; and inputting a to-be-processed first resolution image and the required super-resolution multiple into the training model to obtain a super-resolution image. According to the method, the super-resolution of any scale is realized by using mask filling and image mapping.

Description

technical field [0001] The present invention relates to the field of artificial intelligence, and more specifically refers to a method, system, device and storage medium for super-resolution zooming in on any scale of an image. Background technique [0002] Super Resolution (SR) is the process of recovering high resolution (HR) images from a given low resolution (LR) image, which is a classic application of computer vision. Reconstructing corresponding high-resolution images from low-resolution images (that is, increasing resolution) is important in the fields of monitoring equipment, satellite image remote sensing, digital high-definition, microscopic imaging, video coding communication, video restoration, and medical imaging. application value. [0003] One of the existing technologies, such as ESPCNN, EDSR, SRGAN, RDN and RCAN, only considers the super-resolution of certain integer scale factors (X2, X3, X4) at most. The common point of these methods is that they need to...

AI Technical Summary

Problems solved by technology

Unfortunately, this only scales the image with a fixed integer scale factor

Moreover, the model needs the image to have the same size during the training process, so the original data needs to be resized (redefined size) to grow into an image with the same width, and finally the model is only applicable to the integer scale factor, that is, a model can only be enlarged in one way multiple super-resolution

These disadvantages limit their use in real-world scenarios

Images