Research On Image Restoration Algorithm Based On Deep Learning

With the continuous developing of the information age,the information that people received in daily life is becoming more and more complicated.Among all kinds of information,image information is undoubtedly the most important subject.However,the quality of image information can be affected by defocusing of the imaging system,equipment movement or equipment.The influence of inherent defects often fails to meet the ideal requirements,and thus image restoration technology to improve the quality of image information comes into being.Image denoising is the most basic and important part of image restoration.The quality of denoising results directly affects the success or failure of various subsequent tasks of machine vision.Therefore,it is of great practical significance to study image denoising algorithms.In recent years,thanks to the continuous development of computer computing capabilities,deep learning has got many successes in the field of computer vision through large-scale data sets and high-level deep neural networks,and has gradually become the hottest research direction in the field of machine vision.In the field of image restoration research,the research of traditional filtering methods has reached a certain bottleneck,so many researchers have begun to focus on applying deep learning into image denoising.This paper is based on the theoretical method of image restoration algorithm and the kernel prediction neural network model(KPN),focusing on the research of deep learning methods in image denoising.The paper mainly contains the following work content.(1)In view of the actual limited computing resources,KPN is used as the main network for image denoising,and the experiment is compared with the traditional denoising algorithms,which proves that this method has greater advantages compared with the traditional method.At the same time,it is found that this method exists problems about the high temporal and spatial complexity and denoising effect still having room for improvement.(2)Based on the neural architecture search method(NAS)in automatic machine learning,I analyzed the model's impact of different search spaces,search strategies,and performance evaluation on the performance.Then the network search framework is determined to search the Open image dataset self-made to fined the best architecture of the kernel prediction neural network.Finally,the NAS-KPN model architecture with fewer network layers and strong denoising effect was searched out,and new network parameters were obtained through secondary training on the data set.In the comparison of the results,it was found that the time and space complexity of the NAS-KPN model was reduced and the denoising effect was Improved,but there is still too much noise in the non-pure color area of the result image.(3)Combining the idea of non-local mean(NLM),I add a noise image non-local averaging operation to the image input layer of NAS-KPN,and set up a comparative experiment to explore the performance of image noise information in the image part and the value of the noise level directly,the experimental results prove that the noise information is introduced together at the image and numerical level to achieve the best denoising effect.In addition,in order to verify the generalization ability of the model,comparative experiments were carried out on the real-world image denoising datasets Poly U and SIDD(Smartphone Image Denoising Dataset)of different construction types to verify the generalization ability of NAS and NLM operations on the network model.In summary,this paper uses KPN as the main network,reconstructs the model through NAS,and further optimizes the new model in combination with the idea of non-local mean.The final network model is all in denoising effect,generalization ability,and time and space complexity.A good effect has been achieved.