Research On Meta-transfer Learning For Multi-scale Image Super-resolution Algorithm

Image super-resolution reconstruction refers to the process of obtaining useful information from low-resolution images to generate high-resolution images with rich details and rich content.High-resolution images can not only provide valuable information for tasks such as remote sensing imaging,intelligent security,image retrieval,and medical analysis,but also provide basic framework support for other computer vision tasks such as image classification and small object detection.At this stage,most image super-resolution methods use models that can only generate super-resolution images with a single scaling factor.All of the studies are image super-resolution tasks based on a single scaling factor.By constructing a network model with more parameters,they can achieve better results on single-scale image super-resolution generation tasks.In order to generate multi-scale superresolution images,multiple models must be needed.Few models can simultaneously achieve the task of generating super-resolution images of multiple scales.Using one or several models and generating multi-scale super-resolution images within a limited calculation time is still a difficult problem to solve.Therefore,in response to this problem,this paper proposes a model-agnostic multi-scale image super-resolution algorithm.This method adopts a meta-learning idea that is agnostic of the model.During the training process,images of different scales are sent to the network for training,and the impact of image super-resolution tasks of different scales on the model is balanced,so that a model can optimize different tasks at the same time to achieve the purpose of generating better multi-scale super-resolution images.The model trained with this method can be fine-tuned on specific scales,which can be better and faster for super-resolution image generation at each scale.This paper proposes a model-independent multi-scale image super-resolution algorithm based on one-step gradient.This method only uses one-step gradient information optimization model,which simplifies the division of the data set and reduces the super-parameters while achieving similar super-resolution image generation results.In this paper,a lightweight residual module is introduced,and a lightweight multi-scale super-resolution network is designed.This network has fewer parameters and calculations,which can reduce the generation time of multi-scale super-resolution images.Experimental results show that the multi-scale super-resolution images generated by this method have good results.The quantities of parameters and calculations of the model adopting the lightweight residual module are less than the basic model,and the performance loss is within an acceptable range.