Research And Implementation Of Image Super Resolution Based On Deep Learning

Currently,image is used as one of the most important medium of information storage and communication.The image quality in user generated content is uncontrollable,while people's expectation of high-resolution image should to be met.The task of image super-resolution aims to improve the visual effect of the image,reconstruct high-resolution image from the low-resolution one.According the research in domestic and foreign,the image super-resolution technology based on deep learning has more advantages than interpolation algorithms,which represent traditional image super-resolution technology.However,in the research in recent years,in order to pursue the improvement of networks performance,the image super-resolution networks became deeper and deeper.Because of the high parameters and calculations,some deep networks require high-performance computing devices,which is hinders them from practical usage,especially the usage on mobile devices.In view of this situation,the thesis designs a lightweight image superresolution network,which can apply offline on the mobile devices.The main work completed by the thesis is as follows:(1)In view of the problem that parameters and calculations of current image superresolution networks are too high,the thesis designs a lightweight image superresolution network structure according to the actual application requirements.Besides of standard convolution layers,shuttle-shaped residual blocks also are the basic components of the network.In order to pursue the balance of performance and efficiency,the thesis replaces the standard convolution of the middle layer in the shuttle-shaped residual block with depthwise separable convolution.The PRe LU function is used as the activation function in the network.The sub-pixel convolutional layer is used as up-sampler in the network.And multi-stride branches are introduced into the network to gradually up-sample.The thesis uses mainstream performance evaluation methods to experiment and evaluate the network performance,which proves the effectiveness of the algorithm.(2)The thesis optimizes the lightweight image super-resolution model to be more suitable for real scenes,and develops the SDK jar package to migrate the inference process of the lightweight image super-resolution model to the mobile devices.The implementation and verification of the mobile terminal image super-resolution was carried out on the specific mobile device,and the low-resolution image reading,superresolution image reconstruction and rendering were completed on the smartphone.The actual test results prove that the lightweight image super-resolution model proposed by the thesis can be deployed on the mobile devices to complete the image super-resolution task offline,and can still achieve good results under the condition of limited computing resources and memory resources.The thesis designs a lightweight image super-resolution convolutional neural network according to the actual application requirements,streamlines the parameters and calculations of the network,and achieves a balance between performance and efficiency.The image super-resolution model is deployed on mobile devices and completes inference process using the computing resources of the mobile phone,so that the image super-resolution processing is realized offline.