Research On Image Denoising Based On Deep Learning

Image is the basis of human vision,the objective reflection of natural scenery,and the important source of human understanding of the world and human beings themselves.Images are favored by human because of their visual image,easy to understand,and large amount of information.However,with the more and more frequent use of images,the more and more high-quality image is required,because high-quality images can not only more accurately express the useful information that people want to express,but also provide more accurate information for the subsequent processing of images.In this paper,we studied the tasks of image denoising,image noise level estimation and vehicle logo recognition,the main work is as follows.1.Image denoising is a classical problem in image processing.Aiming at this problem,we proposes an image denoising model based on chi-square distribution and convolutional neural network(CNN),this method take a step forward in cropping image blocks,cropping image blocks based on image noise levels.Firstly,the chi-square distribution is used to estimate the real noise level variance through the sample noise level variance,and then image blocks of different sizes are cropped from the image according to different sample noise level variance.Secondly,a large number of experiments prove that Zhang's Dn CNN has good denoising performance,we improve the original Dn CNN model to improve the denoising performance.Experimental results demonstrate the proposed method achieves superior denoising performance in general image denoising tasks.2.Noise level is an important parameter for many image processing applications.Aiming at the accuracy of noise estimation,we propose an image noise level estimation method based on CNN.Different from previous image noise level estimation models,this model uses the whole image to estimate the noise level.Firstly,the original image is copied and clipped,and different levels of noise are added to form the training data set.Secondly,the noisy images are input into the CNN and the feature map is output in the last convolution layer.Finally,the input noise image is subtracted from the output feature map to obtain a noise map,and then the noise map is input into the full connected layer to obtain the noise level estimation.Experimental results demonstrate the proposed method shows good estimation results,especially when estimating the high noise image level,it has high accuracy.3.Vehicle logo plays a very important auxiliary role in intelligent transportation system.Aiming at the accuracy of vehicle logo recognition,we propose a vehicle logo recognition algorithm based on improved residual network.Firstly,the noisy images are obtained by adding noise to the vehicle logo images,and then the noisy images are processed by applying the image denoising model to obtain the denoised images.Secondly,the residual network is improved by using the improved Leaky Re LU activation function instead of the original activation function.Finally,the structure of the traditional residual network is adjusted.Batch normalization and activation function are placed before the convolutional layer,and reduce the network parameters.Experimental results demonstrate the recognition accuracy of the improved residual network model is 99.8%,which has strong practical significance.