Research And Implementation Of Image Super-resolution Algorithm Based On Neural Network

With the rapid development of computer technology nowadays,images are one of the most intuitive ways to obtain information,and they play an important role in the fields of face recognition,medical imaging,video surveillance,and remote sensing imaging.In the process of image imaging,due to the limitation of lighting conditions and hardware equipment,the details of the image are very easy to cause information loss during the transmission and storage process,which in turn leads to the problem of low resolution of the image.The high-resolution image has rich details and high pixel density,which can meet the actual application needs of image understanding and analysis.Image super resolution technology(Super Resolution,SR)reconstructs a high resolution(HR)image from a low resolution(LR)image through a certain algorithm.For the acquisition of high-resolution images,the most direct way is to use high-resolution pixel industrial cameras.However,in real scenes,due to factors such as production technology and engineering costs,it is impossible to use high-resolution images in many production environments or applications.Resolution camera to collect image information.Therefore,the use of SR technology to obtain HR images is one of the effective methods to solve the above problems.After years of exploration,image super-resolution reconstruction technology is still an important research task in the computer field.In fact,the task is to solve ill-posed problems.First,for any given LR image,there may be different HR images corresponding to it,with slight changes in camera angle,color,brightness,and other variables.At the same time,down-sampling of different HR images may also produce similar LR images,and it will also cause a problem that one LR image corresponds to multiple HR images.Therefore,there is fundamental uncertainty between LR and HR data.In order to reconstruct high-quality HR images,this thesis focuses on the problems of insufficient colors,insufficient details,and unclear edges generated in image super-resolution reconstruction.By analyzing the advantages and disadvantages of existing image super-resolution reconstruction algorithms,explore how to reconstruct HR images that meet actual needs and propose improved algorithms to solve the problems of classic algorithms.At the same time,considering the actual system requirements,an image super-resolution reconstruction algorithm system was designed and developed.The research content of this article is as follows:(1)Aiming at the classic algorithm SRCNN for training and learning only on the Y channel,resulting in the problem of missing colors and details in the reconstructed high-resolution image,a multi-channel R\G\B training and learning method is proposed.Three channels of B are trained independently to extract more detailed image color information,and finally the three-channel network output results are synthesized to obtain HR images with more detailed information.Comparative experiments show that the image reconstructed by the method proposed in this paper is better than the previous algorithm in terms of sharpness and color characteristics.(2)In order to reduce the noise of the super-resolution reconstructed image and obtain more global information,this paper proposes a super-resolution reconstruction algorithm based on wavelet fusion and SRFeat.Using the wavelet transform fusion method for reference,add a branch network on the basis of the original SRFeat network to extract more in-depth feature information.From the low-resolution image,use the wavelet transform fusion method to combine the output of the original SRFeat network with the output of the feature extraction network.Feature fusion makes the detail information of the reconstructed high-resolution image more perfect.After experimental comparison,the improved algorithm is better than the original SRFeat network,and is more in line with the human visual characteristics.It not only reduces the noise generated in the image reconstruction process,but also obtains more global information.(3)Designed and implemented an image super-resolution reconstruction system.The system is divided into three parts:the image to be processed uses the bicubic interpolation method to perform different factor down-sampling preprocessing to obtain the LR image;the two methods proposed in this article are used to super-resolution reconstruction of the LR image to obtain the corresponding HR image;The processed image is displayed on the interface and saved.