Research On Image Denoising Algorithm And Its Performance Optimization Based On U-Net

Today,people are living in an era of flat information,and various screen media are full of people's lives.People's lives are full of various visual media.Counting all the information we usually receive,most of them are visual information,and the image information is an important component.Therefore,how to improve the quality of images has always been a problem that people are constantly pursuing.In addition,with the rapid development of computer vision in recent years,various vision-based algorithms are increasingly relying on image information as basic data.Therefore,many fields such as computer vision have also put forward higher requirements for image quality.However,during the process of image acquisition and transmission,the image will be contaminated by noise due to the complex environment,sensor limitations,imaging principles,natural noise pollution,etc.when the image is acquired,relevant denoising processing is required before using it.Image noise is usually caused by the inherent problems of the image acquisition system and the environment during image acquisition,such as lighting,jitter,and cosmic background radiation.The generation of image noise is a systemic problem,which is extremely difficult to completely avoid.This paper describes several commonly used image noise models,and introduces some traditional image denoising methods based on spatial domain,frequency domain,and mixed domain.However,the field of image denoising is very complicated.This paper selected a few small points for research.Considering the actual performance of various methods and the rapid development of deep learning algorithms in recent years,this paper chose to study image denoising algorithms based on deep learning.Considering the effectiveness of the algorithm and limited computing resources,this paper uses U-Net as the backbone network for image denoising.In the first part,This paper explored the possibility of U-Net for image denoising.Compared with traditional denoising algorithms through experiments,it proves that this method is better than traditional algorithms.However,it was found in experiments that this algorithm for image denoising using deep learning methods also has its shortcomings.The space complexity and time complexity of the algorithm are too large.To solve this problem,this paper optimized the performance of the algorithm mentioned above by applying a densely connected U-Net.This type of network has different depth levels.Through a strategy of network pruning after training,it can effectively improve its runtime efficiency and improve its required storage space.Furthermore,increase the possibility that it can be applied to embedded systems.The generation of image noise in nature is a very complicated problem,and image noise therefore exhibits great differences and uncertainties.This situation is manifested in the problem of denoising: a denoising algorithm can work well for a type of image noise,but it always behaves unsatisfactorily for other types of image noise.In order to solve this problem,this paper proposed a method to optimize the generalization of the denoising algorithm.By applying GAN to simulate the real noise distribution,try to make the neural network fit the real noise distribution.After the simulation of real noise,this paper analyzes the difference between the distribution of simulated real noise and artificial noise,and further shows that the quality of noise generated by deep learning is better than artificial noise based on mathematical models.Finally,based on this real noise model,the generalization of the image denoising algorithm is enhanced.