Research On Single Image Blind Deblurring Algorithm Based On Generative Adversarial Networks

Image blind deblurring algorithm is to deblur the blur image when the blur kernel is unknown.The traditional deblurring algorithm generally is uesd to predict the blur kernel and then deconvolves the blur image with the kernel to obtain the clear image.However,these algorithms have the problems of insufficient generalization and sensitivity to noise,which makes the reconstructed image have a serious ringing effect problem.The image deblurring algorithms combined with deep learning were only used to estimate the blur kernel at first.These algorithms also have the above problems.Later,with the proposal of the encoder-decoder constructure,a deep learning network that directly fits a clear image from a blur image appears.However,such algorithms still have the problems of insufficient model generalization and poor processing of image reconstruction details.Therefore,in this thesis,the conversion idea is to convert the image deblurring problem into an image translation problem,which make it become the problem of converting the image from the blur image domain to the clear image domain.This algorithm fits the clear image data distribution,so it can obtain better network generalization and reconstruction performance.Since there are objects of different scales usually in the blur images,how to balance large-scale information with small-scale detail information becomes the key to the problem.Considering the above problems,this thesis proposes an image deblurring algorithm based on multi-scale residual Generative Advarsarial Networks.First of all,this thesis proposes a new residual network structure that fine-grained residual module,which can simultaneously integrate large-scale information and small-scale information without changing the parameters,thereby improving the quality of blur image reconstruction.Keep its detailed texture information.Secondly,this thesis uses Conditional Generative Adversarial Networks as the basic structure of image translation,which combines adversarial loss and perceptual loss to ensure the content consistency of the reconstructed image better and improve the reconstruction effect.Finally,this thesis uses nearest neighbor interpolation combined with transposed convolution.to upsample the image to avoid the checkerboard effect to the greatest extent.Experiments show that the algorithm proposed in this thesis performs better than the existing image deblurring algorithms on public data sets.Although the improved deblurring algorithm has made some progress,it still has the problems of large parameters and low real-time performance,and the reconstruction image does not work well in some scenarios,and the model generalization is not enough.This thesis improves on the basis of the above algorithms and proposes a deblurring algorithm for Generative Adversarial Networks based on dual discriminators.The main contributions are 1)improved the implementation algorithm of the convolution kernel in the feature extraction stage,and the low-rank decomposition of the original convolution kernel into global convolution makes the network Can increase the receptive field while reducing the parameters;2)Use a double discriminator structure to constrain the generation network.One discriminator pays more attention to the distribution of generated image data,and the other discriminator pays more attention to the distribution of real and clear image data,which can improve the network's generalization.Visualization and reconstruction of visual effects.Experiments show that the improved network has further improved subjective visual effects and objective evaluation indicators.