Research On Image Deblurring Technology Based On Convolutional Neural Network

Images play an important role in human society and are one of the indispensable media for transmitting the information.It will be blurred in acquiring,storing,compressing,and using the image due to the influence of equipment,environment,and noise.Image deblurring removes blurring from the image and restores the image’s original information as much as possible.The deep learning algorithm solves the fuzzy problem faster and better than the traditional filter denoising method.In recent years,many scholars have focused on designing more modules and more complex networks,making the de-fuzzification algorithm redundant and inefficient,which is not conducive to the practical application of the fuzzy algorithm.Therefore,based on the existing research,two deblurring image algorithms are designed in this thesis.(1)A scale cycle deblurring algorithm based on channel selection and cross-layer information transfer.The algorithm mainly contains two aspects: designing feature-based inter-layer information filtering links and multi-scale structure similarity loss functions.Based on cyclic multi-scale architecture,the algorithm uses improved channel attention and interlayer connectivity to optimize the filtering of multi-scale features,designs long network connections to introduce residual learning to help network training,and provides more original features for the output image.At the same time,based on the visual characteristics of the human eye,this thesis designs a multi-scale structure similarity loss and uses the mean square error and structure similarity to achieve different scales of supervision training.Compared with the classical deblurring algorithms in recent years,this algorithm has achieved excellent deblurring results.(2)An asymmetric lightweight image deblurring algorithm based on feature enhancement.A hybrid multi-scale convolution module is designed based on the cyclic multi-scale architecture.The parallel multi-scale structure is used to extract the multi-scale features of the image.The residual pyramid module is also designed,which uses the pyramid structure and expanded convolution to extract multi-scale features while reducing the parameters of the network.The algorithm uses the asymmetric idea and focuses on the design of the network encoder and decoder to achieve the effect of feature enhancement.The decoder achieves feature enhancement by skipping connections and channel filtering.Compared with the classical deblurring algorithms in recent years,the algorithm achieves excellent deblurring effect and faster deblurring speed.In addition,this thesis enumerates the application of image deblurring algorithms in the engineering field to prove that the improved image deblurring algorithm can be applied in practice.