Optimization And Acceleration Of Deep Network For Image Super-resolution

With the continuous development of the image collecting equipment,such as mobile phone,CCD and CMOS camera,the demand of high-resolution(HR)images is still increasing by the pursuit of high quality visual experience.Compared to improve the performance of camera sensors,such as CCD or CMOS,the super-resolution(SR)algorithm provides a more effective approach to solve this demand.In fact,single image super-resolution(SISR)task is a low-level taskn in computer vision.Nowadays,most of CNN-based models devote to train parameters by fiting HR images as much as possible.Three models based on convolution neural network(CNN)are adjusted in this paper.Reduce the network layer that has no effect on the performance,remove the redundant structure in the network,realize the acceleration of different types of super-resolution networks,reduce the operation time and the storage space of the model under the same hardware configuration,and improve the application of the model.It can effectively solve the problem of speed and storage of single frame image superresolution in industrial practice.In order to improve the SISR algorithm,this paper takes three steps.(1)In order to predefine the up-sampling model,it is necessary to preprocess the low-resolution image and upsample it to match the input and output image size,which also leads to the use of high-size images for calculation in the network.Multiply the amount of computation.In this paper,a scheme is designed,which can not only increase the receptive field,but also effectively reduce the amount of computation.At the same time,the structure model can also be effectively applied to the task of image artifact removal,and a clear restoration image can be obtained.(2)For the model based on residual block,the residual block is used to realize jump link,and the batch normalization(BN)method is used to accelerate the convergence of network training.In this paper,the network model is improved and optimized to ensure the accuracy of the model and effectively reduce the time of image processing.(3)At the end of this paper,an optimization scheme based on channel attention model is proposed,which proves that a large number of channel attention structures in the model do not improve the result of image super-resolution obviously.At the same time,a new network layer is proposed to replace the channel attention structure to optimize the calculation speed of the model.On the premise of ensuring that the quantitative index is roughly the same,the operation speed can be improved to 2 to 3 times of the original model,which verifies that the model proposed in this paper is more practical than the original model.