Research On Image Denoising Method Of Tibetan Layout Image

As a long-time information carrier,images have played a very important role in human life and work since ancient times.With the rapid development of high technology and the popularization of electronic digital products,more and more images are used in people’s production and life.However,the image may be inevitably interfered by noise of different intensities due to human factors or physical factors during the acquisition and transmission process,which is not conducive to the processing and use of later images.Therefore,image denoising has become an extremely important branch in the field of image research.Tibetan text,as a written communication tool for the Tibetan people,is different from Chinese characters and other common images.Due to the complexity of the Tibetan text structure,the current research on Tibetan images is not very developed,especially in the Tibetan denoising.Based on the denoising methods of Chinese characters and common images and previous research,this dissertation uses BM3 D and Dn CNN two noise models to carry out the following aspects of Tibetan denoising methods:1.Early data processing.The data set used in the experiment is a Tibetan layout image.In order to better learn and extract the features of the image,the experimental training process requires a large amount of training data.By slicing,the larger layout is cropped into 256?256 size images in order to meet the hardware requirements and improve the training speed,and to achieve the purpose of expanding the data set.Graying the sliced image can reduce the computational complexity and shorten the training time.Add noise to the grayscale image of the Tibetan layout,and add different levels of pepper and salt noise as input to the network.2.A Tibetan layout denoising method based on BM3 D is used.Input the noisy image to the BM3 D model.The processing process is mainly divided into two steps.The first step is basic estimation,grouping by the similarity between the blocks,first performing collaborative filtering,then performing 3D transformation on the three-dimensional matrix,and finally,the aggregation of the overlapping blocks to obtain the basis of the image estimate.The second step uses the basic estimated image obtained in the first step to perform a second estimation and block matching on each block,perform Wiener filtering on the two three-dimensional matrices formed after the matching,and finally obtain a weighted average of the estimates of overlapping blocks Finally,it is estimated that the denoised image is output,combined with the peak signal-to-noise ratio of the denoised image.The experimental results show that the model is good for denoising on the Tibetan version.3.A Dn CNN Tibetan denoising method based on convolutional neural network is used.Aiming at the layout of the Tibetan literature,add pepper and salt noise of different intensities to the Tibetan layout data set.Select the appropriate network structure during the training process,such as using a 3?3 size convolution kernel,using a 17-layer network depth,and training Select the network model with the best denoising effect times,and continuously optimize to obtain the training results with the best denoising effect in the Tibetan layout.After the training is completed,the test set is input into the network,and the denoising effect of the model under different noise intensities is analyzed according to the denoising results of the Tibetan layout test set.Experimental results show that the Dn CNN denoising algorithm has a good denoising effect on the Tibetan layout.Comparing the peak signal-to-noise ratio of the two models after denoising on the Tibetan version,it is concluded that the Dn CNN denoising method can achieve a better denoising effect on the Tibetan version than the BM3 D denoising method.