Research On Image Restoration Algorithm Based On Deep Learning

Digital images are an important way for people to obtain information.However,digital images are often subject to the limitations of the external environment or their own hardware technology in the acquisition and transmission,making the acquired images have noise,rain and other interference.Image recovery has been a hot research topic in image processing,computer vision and related applications.In recent years,many deep learning-based image recovery algorithms have been proposed and have achieved excellent results in various performance indexes,but deep learning relies on a large number of data samples and there is still the problem of noise interference residue in the image recovery process.For this reason,this paper mainly focuses on noise,rain lines,raindrops and other interference factors,and researches image recovery algorithms based on deep convolutional neural networks,with specific research contents including:(1)A pairwise residual denoising network combining local and non-local attention is proposed.The network adopts a pairwise residual network learning structure and introduces convolutional kernels of different sizes to improve the feature extraction ability.In addition,this paper adopts a hybrid attention mechanism,which enables the convolutional neural network to pay attention not only to the local features of the image but also to portray the long-distance dependencies in the image.The experimental results show that the algorithm can effectively suppress multiple intensities of noise and has better removal performance for high-intensity noise.Compared with common denoising algorithms,the denoised images can retain more structural information and achieve excellent performance in all indexes.(2)A raindrop dataset was constructed to simulate a real environment.An image acquisition platform was first built using cameras and glass to simulate outdoor video surveillance equipment.The situation of raindrops falling on the glass is simulated by using manual sprinkling form to acquire paired images with/without raindrops.And the acquired data are manually screened,image cropping and image pre-processing sessions to obtain 603 pairs of raindrop image datasets.The verification of the relevant experimental results proves that this dataset can effectively reflect the characteristics of the images captured by the outdoor vision system in rainy days,and can provide a realistic raindrop image dataset for the research and application of image derain.(3)An image derain network based on a dynamic attention allocation mechanism is proposed.In order to better capture rain lines of different sizes,densities and orientations in the de-rain task,an adaptive method of picking the position of attention modules is proposed.In addition,by combining with a multi-stage feature extraction structure,the information extraction capability is effectively improved to provide rain line information for the subsequent recovery process.After testing with seven deep rain removal algorithms on both public and real rain image datasets,it is demonstrated that this algorithm is not only effective in removing synthetic rain lines,but also in suppressing the interference of real rain lines as well.