Research On Image Super Resolution Reconstruction Based On Convolutional Neural Network

Image quality plays a vital role in various image tasks.To some extent,it determines the difficulty level of the task and the effect of the completion.Restoring image quality with super-resolution technology has become a research hotspot,but the super-resolution reconstruction task is a morbid problem,because you want to recover higher resolution images from low resolution images.Changing hardware devices can lead to increased image acquisition costs.If the medical image acquisition and replacement device may also affect the patient's health,it is the best choice to improve the image resolution from the software point of view,that is,recover high-resolution images by super-resolution reconstruction.In this thesis,further research on the defects of existing super-resolution reconstruction algorithms based on deep learning,and a series of innovations have been achieved.The work in this thesis mainly includes the following two aspects:(1)Super-resolution reconstruction of two-dimensional natural imagesFirstly,for the super-resolution reconstruction task of two-dimensional natural images,a network structure is proposed to facilitate feature reuse.By improving the utilization of features,the effect of reconstructing images is further improved.An improved jump layer structure is added to the feature reuse structure to make the features better fused.Secondly,for the two-dimensional super-resolution reconstruction task,a two-parameter loss function is designed.By changing the two adjustable parameters,select the loss function that best fits the network structure of this thesis.Comparison by experiment,the feature reuse structure proposed in this thesis is better than the similar algorithm in the two-dimensional natural image super-resolution reconstruction.In the other four similar public datasets,the proposed structure has significant advantages,which proves that the proposed algorithm has strong generalization ability.(2)Super-resolution reconstruction of 3D medical imagesAiming at the super-resolution reconstruction of 3D medical images,a network structure based on cavity convolution is proposed.Without increasing network parameters,the model's receptive field is increased to better extract features.Secondly,by adding a layer jump connection structure,the shallow local information is further integrated with the deep global information to improve the reconstructed image quality.In three-dimensional medical image super-resolution reconstruction task,the proposed super-resolution reconstruction algorithm based on cavity convolution has significant advantages over other similar algorithms,and can reconstruct higher quality images.