Research On Image Deblurring Algorithms Based On Generative Adversarial Networks

With the development of modern imaging technology,images can record and transmit the abundant information in various scenes.However,the captured images often suffer from motion blur caused by factors such as lens jitter or object motion.This not only affects people’s visual perception,but also poses great difficulties for subsequent image analysis and processing.Deep learning algorithms have become increasingly popular for image deblurring in recent years.However,these algorithms fail to fully utilize the global dependencies among structural features of the image,leading to limited performance.To address this issue,this dissertation focuses on the research of image deblurring algorithm based on generative adversarial networks.The main research contents of this dissertation are as follows.Firstly,an attention generative adversarial network is proposed to address the issue of motion blur in captured images.The proposed network adopts the dense residual modules and cascaded criss-cross attention modules in the generator,allowing for adaptive integration of feature information and global contextual relations for more effective handling of complex motion blur.Additionally,a Markov discriminator with gradient normalization is utilized for discriminating sharp and deblurred image pairs on local patches,improving texture features of restored images.On this basis,in order to make the contours and details generated by the network not distorted,a multi-component loss function is adopted to improve the authenticity of the restored image in terms of color and texture.Secondly,to address the issue of large number of model parameters in the proposed attention generative adversarial network,an improved lightweight network based on multi-scale discrimination is proposed.This improved network replaces the 7x7 convolutions and regular residual modules in the original generator with depth-wise separable convolutions and inverted residual modules with linear bottlenecks,respectively.This results in a more lightweight network.Furthermore,the multi-scale discriminator is used to discriminate the input images across multiple scales with varying receptive fields.Therefore,the multi-scale discriminator uses the feature information contained in the image at different scales to better guide the generator to restore blurred objects at different scales.Finally,qualitative and quantitative comparative experiments were conducted on public datasets.The effectiveness of the proposed algorithm is verified by comparative experiments.