| In modern information society,image is an important means of storing and disseminating information,which can make people intuitively perceive information.In modern people’s work and life,image is one of the indispensable ways to perceive the world information.With the improvement of the quality of work and life,people have improved the standard of image quality.The emergence of image super-resolution technology can improve image quality under the condition of limited equipment,and meet the needs of remote sensing imaging,medical diagnosis,astronomical observation and many other fields.Therefore,people pay more and more attention to the development of this field.In recent years,the continuous breakthrough of computer technology has promoted the rapid growth of depth learning.Especially in the field of computer vision,deep learning has broken through many visual problems.The super-resolution method based on deep learning has outstanding effect,surpasses the traditional method,and is the key research direction in this field.However,due to the ill-posed problem of image super-resolution,many current methods have not reached a satisfactory level.In order to effectively to reconstruct highresolution images,this paper deeply studies the ill-posed nature of image super-resolution from the perspective of calculation and image quality.In this work proposed an image superresolution reconstruction algorithm based on improved Invertible Rescaling Net(IRN).The feature extraction module is designed by using the dense residual structure combined with the attention mechanism to solve the uncertainty caused by the deep network and allocate more attention to learn the high-frequency information of the image.In the feature extraction module,dilated convolution is introduced to improve the receptive field and keep the parameters unchanged,which reduces the parameters of the whole network.Using wavelet transform high-frequency subbands interpolation to design the potential variable of IRN network,solve the problem of potential variable storing image high-frequency information,and improve the ability of network to capture image high-frequency information.A superresolution reconstruction system is designed by using the improved algorithm in this paper.The system is constructed based on Web and can realize module functions such as superresolution image reconstruction,model training and so on.Comparative experiments with advanced methods are carried out on four public benchmark data sets to verify the effect of this algorithm.Experimental results show that the proposed algorithm can reconstruct high-quality images,and the complexity of the network model is lower,so it has a certain application value. |
