Gaussian Mixture Model Guided Dictionary Learning Method For Single Image Deraining

Under rainy days,the images collected by outdoor vision system are seriously degraded by the scattering and blurring of raindrops,which including blurring,information occlusion and contrast reduction.This severely restricts the correct understanding and decision making of the environment by outdoor computer vision systems.Therefore,image deraining is helpful to improve the accuracy and reliability of various computer vision algorithms.Video-based deraining method can take advantage of multiple frames of information,and significant progress has been made.In contrast,it is more difficult for single image deraining owing to the lack of temporal correlation and context information in multiple adjacent frames.This paper focuses on the in-depth research and analysis on the single image deraining problem.The main work is as follows:(1)In order to combine the prior knowledge of natural image and rainy image,this paper proposes Gaussian mixture model(GMM)guided low-rank dictionary learning algorithm and applies it to a single image deraining.Firstly,the model exploits the GMM to learn the external dictionary from a large number of natural image patches.Secondly,we exploit the learned external dictionary to guide internal dictionary learning.Meanwhile,internal dictionary with low-rank constraint is incorporated into the objective function of dictionary learning.By exploring GMM learn the external dictionary from a large number of natural images,which helps to effectively recover sparse details in images,while learning internal dictionaries from rainy images is conducive to the recovery of dense details in images.Through the replacement of the external dictionary to continuously guide the internal dictionary learning,the prior knowledge of natural images and rainy images can be more organically integrated,thereby improving the performance of the image deraining.Experimental results on both synthetic and real-world rainy images demonstrate the effectiveness of the proposed method.(2)In view of the above-mentioned Gaussian mixture model(GMM)guided lowrank dictionary learning algorithm neglect the deviation between the training sample and the test sample,which decrease the accuracy of the coding coefficients during reconstruction and the generalization performance of the image deraining model.In this paper,guided dictionary learning algorithm with group sparse residual constraints is proposed for single image deraining efficiently.This model introduces residual constraint based on guided dictionary learning method,which can effectively improve the reconstruction and generalization ability of learned dictionary,thereby obtaining better image reconstruction performance.Furthermore,based on the criterion of image nonlocal self-similarity,the group structure sparse representation is introduced to ensure that similar image patches have the similar coding coefficients.Compared with other algorithms in the synthetic image and the real image,the experimental demonstrate that the reconstructed image with the proposed algorithm has better highquality and more detailed information,and visual effect can be significantly improved compared with the state-of-the-art other algorithms.At the same time,the generalization of the proposed method is verified on the real rainy image of night scene.(3)Group-based sparse representation take into account the correlation between similar image patches,which can well describe the global characteristics of the image.Meanwhile,patch-based sparse representation can keep the local sparsity of the image well.Considering the complementarity of the two methods,this paper proposes patchgroup sparse joint representation algorithm based on guided dictionary learning,and applies it to single image deraining.The model can organically combine the advantage of patch-based sparse representation and group-based sparse representation to improve the performance of image deraining.The experimental results demonstrate that the proposed method not only has obvious improvement over the existing methods,but also can better preserve image local structure and suppress the artifacts.So the visual effect of derained image is more clear and natural.