| In recent years,with the need for image clarity in fields such as autonomous driving,medical imaging,and road monitoring,visual effects have become increasingly demanding.In order to cope with this phenomenon,there are more and more studies on image deblurring,and the corresponding effect of image deblurring has also been improved to a certain extent.However,in a complex living environment,there can be numerous factors that cause blurry images.For example,the rapid movement of objects,solar radiation,atmospheric turbulence,etc.will cause image quality to deteriorate,which brings many challenges to the study of image deblurring.The traditional image deblurring method in order to portray the characteristics of the original image mostly use prior modeling,this way is required to be manually designed,the ability is poor,once the prior model design is more complex,then the model solution process is very difficult.With the application of deep learning to image deblurring technology,a large number of data sets can clearly predict image features by learning,and achieve good results in some competitions and practical applications.With the proposed generative adversarial network,it is trained by adversarial learning to achieve good performance in the field of generative images.Since then,generative adversarial networks have also begun to be applied to the field of image deblurring,such as the Gan deblurring network(De Blur Gan)proposed by Kupyn et al.in 2018.Based on this,based on the deep learning method of generative adversarial networks,this thesis conducts the following research:(1)An image deblurring method based on dense residual connection is proposed.Aiming at the defects that the network structure of the existing image deblurring algorithm can not be fully utilized after feature extraction through convolution operation,and the high redundancy in the learning process leads to too low work efficiency,this thesis proposes a deblurring method based on generating antagonism network(Gan),which uses dense residual modules in the main network of the generator.First of all,create a dense connection module with feature reuse.The input of each layer node comes from the features of all previous layers,and the design of each layer network module is particularly narrow,so only a few features are learned to achieve the purpose of reducing redundancy.Then,combining the dense connection module with the residual connection module can further alleviate the problem of gradient disappearance.Finally,when calculating the perception loss of the real image and the generated image,using the pre-activated features for VGG extraction features can improve the performance of the network,and keep the brightness of the reconstructed image consistent with the real image.(2)An image deblurring method based on attention mechanism is proposed.Se Net is introduced into the generator backbone network to make the network pay more attention to features that are more useful to the task to further enhance the model performance;Up-sampling in the generator network uses interpolation instead of deconvolution to suppress the checkerboard effect.(3)In order to make the image deblurring method proposed in this thesis be applied in the actual scene,the multi-functional image deblurring application system can be finally realized through analysis and design,and multiple image deblurring technologies can be integrated in the human-computer interface.By inputting the blurred image to be processed and selecting the deblurring algorithm,the results after deblurring can be displayed intuitively,and the parameters of the deblurred image can be viewed. |
