Super Resolution Reconstruction Algorithm Based On Internal Dataset Learning

In recent years,deep learning has been widely used in super-resolution reconstruction tasks.Most of the data of super-resolution research are based on external examples.This kind of method obtains the super-resolution reconstruction network model by training the nonlinear mapping function between high and low resolution image patches,which has made great progress compared with traditional methods.However,the method based on external instances needs a lot of training data,and can not guarantee that the data set contains all the mapping relationships between high and low resolution image blocks;The complex network model also makes the training process very time-consuming,especially when different scale network models need to be trained;In the real world,the degradation model of low and medium resolution images is often very complex.These images may be interfered by noise,blur and other factors,resulting in the uncertainty of the reconstruction effect of a specific image.Other methods focus on a single image and use internal examples to reconstruct high-resolution images,avoiding the limitations of external examples.On this basis,two super-resolution algorithms are proposed by combining internal data sets with deep learning:(1)A super-resolution reconstruction algorithm based on multi-scale CNN:The algorithm is inspired by zero-short SR(ZSSR)algorithm.ZSSR combines internal data set with deep learning for the first time,which overcomes the defects of external data set.At the same time,it can reconstruct satisfactory high-resolution image in limited time.But ZSSR method also has some defects.Firstly,the network directly performs bicubic interpolation on the input low resolution image,and smoothes the little detail information compared with the external data set,which is not conducive to the training of the network;Secondly,the advantage of internal data set is that it can reconstruct high-resolution image by self similarity in multi-scale range,which ZSSR does not make full use of.Based on these problems,this paper proposes a multi-scale mapping method based on CNN input space and interpolation space.High-resolution images are obtained by single-scale feature mapping module,cross-scale feature mapping module,multi-scale feature fusion and global residual learning module.The network trains the feature mapping from low-resolution image to high-resolution image at the same time,which makes full use of the multi-scale information of internal mapping.Compared with several super-resolution reconstruction algorithms based on internal data sets,the proposed algorithm improves the quality of image reconstruction.(1)A super-resolution reconstruction algorithm based on attention mechanism:The existing super-resolution reconstruction algorithms based on attention mechanism usually include channel attention mechanism and spatial attention mechanism.Among them,channel attention mechanism is widely used.This kind of algorithm optimizes the different weights of each channel through full connection layer operation,and pays more attention to the more important features,so as to achieve the purpose of optimizing the network structure.However,for deep network,with the change of network depth,the features extracted by convolution operation are different.In image super-resolution task,shallow network is easier to extract low-frequency information such as image structure and location,while deep network is responsible for extracting high-frequency information such as image details and edges.Therefore,it is necessary to give different weights to different depth features.This algorithm proposes a deep attention mechanism module(DAB)based on the characteristics of deep network.Firstly,shallow features and deep features are extracted respectively.In the process of network forward propagation,different weights are given to cascaded shallow features and deep features with the help of deep attention module,The learning focuses on the features that have more obvious effect on the super-resolution reconstruction task,which speeds up the convergence speed of the model and improves the quality of image reconstruction.