Research Of Image Denoising Algorithm Based On Convolutional Neural Network

In daily life,images have become the most commonly used information carrier in human social activities.However,digital images are often corrupted with different kinds of noise during acquisition and transmission,which degenerate the quality and performance in later tasks such as image recognition and medical diagnosis.Denoising technologies are designed to reduce or remove the influence of noise,while preserving information in the original images at the same time.Therefor image denoising is necessary before image recognition,analyzing and application.Current spatial denoising algorithms are mostly on the basis of filters such as mean filter or median filter.In frequency domain,denoising algorithms are based on the wavelet transform or Contourlet transform.So far,convolution neural network has been applied to the field of computer vision successfully.However,it is mostly used for pattern recognition such as image classification,identification and target detection.In the field of image processing,convolution neural network is used to improve image resolution,remove blur and so on.As a classic model of neural network,convolutional neural network have made a great contribution to pattern recognition,but little to image denoising.Convolution layer in convolutional neural network is used for image feature extraction;sampling layer reduces the structural parameters of the network.According to this structure characteristics o,we propose an image denoising algorithm based on convolutional neural network.The biggest innovation is that compared with conventional convolution neural network,the proposed network includes only convolutional layer,not sampling layer.while denoising process includes extracting blocks,nonlinear mapping and image reconstruction,which directly maps noise image to clean image.This study expands the applications of convolutional neural network in image processing.This paper mainly used Gaussian white noise images as input,and compared with current similar methods.For different types of noise,we trained different convolutional neural network models which showed better performance compared with similar algorithms.Peak signal to noise ratio is used here to quantitatively assess different denoising algorithms.