Research On Image Motion Deblurring Based On Deep Learning

Image is an important source of environmental information in many research fields such as autonomous driving and unmanned aerial vehicle swarming.However,due to limitations of imaging devices,environments,human factors(e.g.camera shake,relative motion between camera and target)and so on,hardly can image motion blur be avoided,which greatly reduces the information quality and relevant algorithm performance.It is greatly helpful to recover potentially clear image and then enhance the richness and reliability of image for further scientific research.The current deblurring method is mainly based on deep learning,but there still exist problems such as insufficient blur removal and weak adaptability of the algorithm.In this paper,we focus on the paired and unpaired image motion deblurring,and then research the learning-based deblurring method of high efficiency and strong generalization ability,aiming to improve the deblurring quality while reducing the computational complexity and enhancing the adaptability of the algorithm in complex environments.The main work and results of this paper are listed as follows:1.We propose a new multi-scale residual feature extraction model for motion deblurring,which embeds multi-scale residual block with multi-channel feature extraction path.Meanwhile,the feature paths of adjacent scales are added in the channel dimension to ensure better feature extraction and avoid gradually information loss.Compared with the current typical learning-based motion deblurring model,the deblurring results obtained by the feature extraction model have 0.39 d B and 0.0047 improvement in terms of PSNR and SSIM index.2.We propose an end-to-end image deblurring algorithm with multi-scale residual channel attention to directly learning the mapping between the blurry image domain and the clear image domain,which effectively avoids the blur kernel estimation process.The low-frequency content and high-frequency details of the image are separated and differentiated by the residual channel attention module in the algorithm.The module pays attention to processing the high frequency detail information and avoids the redundant calculation of a large amount of low frequency information.Compared with the existing representative deblurring algorithm,the proposed end-to-end image deblurring algorithm for multi-scale residual channel attention has 0.79 d B and 0.0117 improvement in terms of PSNR and SSIM index,with four times improvement in runtime.3.We propose an unpaired image deblurring model based on cycle-consistent generative adversarial network.The model defines the deblurring problem as the mutual conversion between the blurry domain and the clear domain.It consists of two generative adversarial networks,respectively implementing the conversion of the blurry image domain to the clear image domain and the clear image domain to the blurry image domain,which constitute data closed loop.Defining the loss function by constraining the distance between the converted output image and the input image,the model finally achieves unsupervised image motion deblurring.According to the characteristics of the processing task,the training dataset can be constructed specifically,making the model more flexible,efficient and applicable.