Research On Blind Deblurring Method Of Single Image Based On Deep Learning

Image deblurring is a classic computer vision problem,and its purpose is to recover a corresponding clear image from a blurred image.Image deblurring methods can be divided into blind image deblurring and non-blind image deblurring according to whether the blur kernel information is known or not.In this thesis,aiming at the problem of blind deblurring of images and combining deep learning,three deep networks are built for image deblurring.The specific research contents are as follows:First,a blind deblurring method for multi-scale dynamic scenes based on attention mechanism is designed.Since multi-scale information can effectively improve the performance of image deblurring,a multi-input multi-output multi-scale network structure is designed for image deblurring.This method effectively fuses features from different scales by utilizing an attention-based feature fusion module.In addition,a context block with different dilation rates is added to enlarge the receptive field and extract features at different scales.The experimental results show that the method in this chapter is not only beneficial to improve the performance of image deblurring,but also has obvious advantages in terms of speed.Secondly,a blind deblurring method for dynamic scenes based on Fast Fourier Transform is designed.In this method,an encoder-decoder network structure is constructed by using the Residual Fast Fourier Transform Block and the Residual Fast Fourier Transform Block based on half-instance normalization from the perspectives of the frequency domain and the spatial domain.In the encoder stage,the receptive field of each scale is enlarged by using the Residual Fast Fourier Transform Block based on halfinstance normalization.In the decoder stage,by using the Residual Fast Fourier Transform Block to capture global information and local context information,which can better reconstruct high-quality clear images.Experimental results show that this method can significantly improve the quality of restored images.Finally,a blind image deblurring method based on Transformer and wavelet transform is designed.Since the type of blur is not only motion blur,but also defocus blur,so this method uses Transformer module as the basic module and uses wavelet reconstruction module to design a network that can remove motion blur and defocus blur.The wavelet reconstruction module can convert the spatial domain features to the wavelet domain for reconstruction,thereby improving the learning efficiency of the network.