Research On Image Super-resolution Method Based On Improved Convolutional Neural Network

Computer network and internet technology have been more widely used because of the rapid development of information technology.With the advent of the era of big data,people's dependence on information has become increasingly prominent.Furthermore,people put forward higher requirements for information processing and information exchange,including high-resolution image information.Traditional image information mainly saves bandwidth by reducing image resolution,which often makes the image can not meet people's expectations.Moreover,in many cases,high-resolution and super-resolution images are not used for image signal acquisition due to the production process and engineering cost.Therefore,it is particularly important to acquire high-resolution images by super-resolution technology.Image super-resolution refers to the improvement of low-resolution images to high-resolution images through certain algorithms,so that high-resolution images have higher pixel density and better details,which can effectively make up for the deficiency of hardware acquisition information equipment such as sensors.Human cognitive mechanism and expressibility can provide an important theoretical basis for intelligent computing of image processing.The central production system of ACT-R model mainly includes three processing stages: matching,selection and decision making,which correspond to the analysis,construction and image processing in the process of image reconstruction respectively.Therefore,taking ACT-R model of cognitive framework as the background and combining with deep learning method.This paper constructs an ACT-R research model applied in the study of image super-resolution reconstruction in order to improve the image super-resolution reconstruction effect.In the matching stage,the network training is mainly carried out to prepare for image reconstruction.In the selection stage,the recurrent convolutionalneural network(RCNN)model is established.Then the residual image is constructed and reconstructed;In the decision-making stage,the simulation experiment is mainly carried out on the algorithm proposed in this paper,and the algorithm in this paper is evaluated by comparing with other methods.Firstly,we add a circulation layer to the basic convolutional neural network to train an improved end-to-end deep threshold cyclic convolutional neural network.The depth of the convolutional network is increased on the premise of not changing the number of convolutional network layers.Secondly,the neighborhood embedding algorithm is used to compensate the residual part of the image,so that the high resolution image with higher pixel density and better details.Finally,the reconstructed image and the reconstructed residual image are fused to obtain the final reconstructed high-resolution image.