Research On Image Super-resolution Reconstruction Method Based On Deep Learning

In recent years,the development of neural network technology has promoted the development of artificial intelligence.The most representative one is the field of computer vision.This field contains multiple research directions.Other research directions in this field have been based on image superresolution reconstruction.Therefore,a large number of scholars conduct research in this direction.At the same time,the technology also has high requirements in all aspects of real life.For example:monitors are installed in every corner of the town to facilitate the police to track the whereabouts of prisoners;in the medical field,doctors can use film to determine the patient's condition and determine the severity of the condition;and collect remote sensing from high-altitude areas image.Land surveying,etc.However,the current super-resolution reconstruction model has too many parameters,the model is too large,and the quality of detail restoration is not high,which is not convenient for practical application.In response to the above problems,this paper proposes the following two singleimage super-resolution reconstruction methods:(1)Single image super-resolution reconstruction algorithm based on dense Inception.Generally,the simplest and safest way to improve the performance of deep neural networks is to stack convolutional layers in a fully connected manner to increase their depth and width.However,each additional convolutional layer will bring exponential growth of parameters,which requires more computing resources.Under the premise of a small number of training sets,overfitting is prone to occur.For the problems mentioned earlier,this paper constructs a reconstruction model based on dense connections and Inception modules.The model introduces the Inception-Residual Network(Inception-ResNet)structure to extract image features,and converts the convolutional layer previously connected by a direct stacking method into a non-tight connection method.The simplified dense jump connection is used globally,and only when each module is constructed can it be output to the reconstruction layer to avoid redundant data and increase the amount of calculation.On the premise of ensuring the improvement of network performance,make the number of parameters as small as possible.(2)Single image super-resolution reconstruction method based on residual Inception.Many studies have shown that building a wider and deeper network can significantly improve the reconstruction performance of image super-resolution,but this strategy has many parameters and is difficult to train.At the same time,in the reconstruction process,only the top-level features are used for reconstruction,and a lot of detailed information may be lost.In response to the above problems,this paper constructs a single-image super-resolution reconstruction model based on residual Inception.The overall model of the model adopts a method of combining global and local residual connections,fusing the characteristics of each core module for reconstruction,which is more conducive to training deeper networks than dense jump connections.The main structure of the core module adopts an improved version of Inception,which can extract image features at multiple scales,increase the globality of feature information learning,and restore more high-frequency details,which is conducive to reconstruction.