Research On Blur Removal Based Method On Convolutional Neural Network

As an important information carrier in human daily life,the quality of images directly determines whether effective information can be obtained in a timely and accurate manner.In the process of image collection,the image is often blurred due to equipment shake and the shooting object in motion.As a result of the loss of image information,image deblurring,as an important research direction in the field of image processing,has been widely used in transportation,astronomy and military fields.In recent years,with the rapid development of convolutional neural networks and deep learning in the field of image processing,their efficient and fast processing methods have gradually attracted people’s attention,and they have higher accuracy for specific blur problems.Based on the above problems,this paper is based on Convolutional neural networks have done the following researches:1.For the problem of complex and difficult to form motion blur image scenes,based on the multi-scale image in the Gaussian pyramid of the image,combined with multi-scale image information extraction and multi-scale feature extraction to establish an end-to-end deblurring model,adilated convolution module was designed with the purpose of expanding the receptive field and the model depth was broadened based on a small number of parameters.The experimental results show that it can run faster and can handle different sizes on the premise of ensuring the deblurring effect.Image of complex motion blur.2.Aiming at the problem that the neural network model cannot accurately establish the image blur formation process,according to the argument that the convolutional neural network can obtain the statistical information of the image before going through any learning process,constructing the image prior information and blur respectively.Generate a prior information network and constrain the optimization of the image composition through input noise to ensure that the network model can correctly extract prior image information and generate clear images without relying on a large number of training datasets.3.Aiming at the problems of complex operation and time-consuming iteration of traditional algorithms,the traditional deblurring problem is analyzed based on the maximum posterior method.Multi-scale priors are used as a preprocessing process,and the relationship between the blurry image and the blur kernel is derived A neural network generator and optimizer with a residual network as the main body is produced,and relevant datasets are produced to train the blurry image in combination with the blur kernel to improve the training speed and achieve good restoration results.