Research On Single Image Restoration Algorithm Based On Deep Learnin

With the rapid development of deep learning,digital image restoration technology has become a popular research field,aiming to restore degraded images through algorithms.Although great progress has been made in image restoration algorithms in medical fields and other fields,there are still problems such as poor restoration effect and low model stability.Therefore,according to the different factors causing image degradation,this paper carries out the research of image deblurring,image rain removal and rain map generation,and proposes a single image restoration algorithm based on deep learning.The main work contents are as follows:(1)To address the issues of flexibility and effectiveness in improving a single deep learning model for image deblurring tasks,proposed the Dual Control Network-based Single-Image Deblurring Algorithm.For enhancing effectiveness,the algorithm is trained in an end-to-end manner.In terms of flexibility,explicit degradation principles and prior constraints are incorporated.Specifically,the algorithm consists of a control module,degradation branches,and processing branches.The control module predicts hyperparameters based on different types of blur kernels and utilizes them to finely adjust the output of the degradation branches,thereby achieving the degradation principle.Simultaneously,the control module utilizes noise as prior knowledge to predict hyperparameters and constrain the output of the processing branches,thus fulfilling the prior constraint principle.Additionally,a cyclic skip-connection structure is introduced to enhance the mapping of low-level features to higher layers and address the issue of gradient vanishing caused by backpropagation.we use Peak Signal to Noise Ratio(PSNR)and Structural Similarity Index Measure(SSIM)as objective evaluation indexes and visual perception as subjective evaluation indexes.Experimental results demonstrate that the proposed algorithm effectively removes various types of blur artifacts and restores clear images,thereby validating its flexibility and effectiveness in deblurring multiple types of blurry images.(2)To address the problem of ineffective separation between the rain streak layer and the clear background layer in rainy images,a single-image rain removal algorithm based on a multistage recursive network is proposed.The algorithm consists of three stages,each dedicated to extracting the rain streak layer by subtracting it from the input rainy image and recursively obtaining a clear image.To tackle the issue of insufficient extraction of rain streak features at different scales,a scale-expansion convolution module is introduced,combining dilated convolution operations with a multi-scale strategy.To facilitate better information transfer between stages,convolutional gated recurrent units are used for inter-stage connections with parameter sharing.Experimental results demonstrate that the proposed algorithm achieves higher objective evaluation scores on synthetic test sets and provides better subjective visual perception on real-world test sets,retaining more image detail(3)To address the issue of the limited diversity in rain streak distribution in synthesized rain images,a rain image generator algorithm based on generative adversarial networks(GANs)and image deraining research is proposed by taking advantage of the good generation ability of generative adversarial network.The algorithm is trained using two datasets,which effectively tests the ability of the proposed algorithm to generate rain maps in two different backgrounds.The algorithm comprises an encoder,a generator,and a discriminator.The encoder is used to encode the input image features and generate latent space vectors.To enhance the diversity and high-quality of generate rain images,the algorithm utilizes reparameterization techniques to handle the latent vectors.In order to better capture the global correlations of image features and improve the stability of the algorithm,the discriminator introduces self-attention mechanism and spectral normalization techniques.Additionally,in order to verify the influence of the quality of the generated rain image on the performance of the rain removal algorithm,the comparison experiment between the training and testing effect of the generator generated rain map was carried out.The algorithm evaluates its performance using FID scores,loss function curves of the generator and discriminator,and subjective visual evaluation of the generated rain images.The experimental results demonstrate that the algorithm produces diverse and highquality rain images.Moreover,the rain image is generated to participate in the training of the rain removal algorithm,which effectively improves the performance of the rain removal algorithm.