Image Deblurring Based On Convolutional Neural Networks

Camera has been widely used in our daily lives and the image,as a product of the camera,provides an important way to convey information.However,the acquired image can be blurred under some unexpected situations,e.g.,the movement of the object,focusing on false object and lack of light conditions.The degraded image often cannot meet the needs of human beings and even cause serious economic loss.Therefore this thesis studies image deblurring.First of all,the theoretical background of image deblurring based on the convolutional neural networks(CNN)is investigated.Several kinds of popular deblurring models and the principle of CNN model are introduced.Next,the deficiency of existing deep learning based deblurring methods is analyzed and a new CNN model called fast CNN(FCNN)is introduced to deblur the image quickly while keeping the high frequency details.The method first deblurs the image in the Fourier domain with regularizations to obtain a pre-processed latent image.Then,the blocks of the pre-processed results and their corresponding clean blocks are served as the inputs and labels to train the FCNN,so that the mapping function from blurred image to potential clean image is obtained.Consequently,a clear image can be obtained by the trained FCNN.In the pre-process,two regularizations with gradient constraint of Gaussian model and the smoothness constraint are incorporated,which the image priors are used to recover the image directly and get a robust value as a foundation for FCNN model.FCNN model consists of a four-layer convolution and activation function which eliminates the image blur while reducing the time complexity.Qualitative and quantitative experiments demonstrate that the proposed method can adopt the parameter-sharing property of CNN effectively and thus reduce the number of training parameters significantly.It also greatly reduces the computational complexity in comparison with the existing deep learning based algorithms while keeping the image details well.Finally,the improved FCNN,P-FCNN,is proposed for deblurring,which incorporates the least squares filtering algorithm and increases the depth of model,aiming at improving the adaptability of FCNN method.In this model,the advantages of the 1×1 convolution layer are analyzed and then applied to FCNN to obtain the?-FCNN model.The depth of the model is increased,while the number of training parameters is reduced.The least squares filtering algorithm which can enhance the image edge after filtering is also taken.Consequently,the least squares filtering algorithm combined ?-FCNN model deblurring method,?-FCNN,is proposed.The method first deblurs the image using the least squares filtering algorithm to obtain a pre-processed latent image.Then,the same training process of FCNN is applied to obtain the mapping function,so that a clear image can be obtained by the trained?-FCNN.Experiments demonstrate that the advantage of ?-FCNN in adaptability in comparison with other deep learning methods.