Research And Improvement Of The Deep Learning Based Image Denoising Algorithm

In recent years,with the development and popularization of the Internet,"artificial intelligence" has been well known.Since the AI godfather Geoff Hinton proposed a multilayered neural network,the learning and research of artificial neural networks has been widely concerned by relevant scholars,thus accelerating the development of artificial intelligence.With the development of artificial intelligence,the method based on deep learning has achieved outstanding performance,especially in the fields of image recognition,speech recognition,image denoising and image reconstruction.In a real-life scenario,due to imperfections in equipment and systems or the presence of noise in the image captured in a low-light environment,the image is subject to additional noise during compression and transmission.Noise can reduce the sensory effect of the image while affecting the compression performance of the image and the performance of computer vision processing.This paper focuses on the basic principles of image denoising and image denoising based on non-local self-similarity principle and deep learning techniques.Block Matching 3 Dimensional(BM3D)and Denoising algorithm based on Convolutional Neural Network(DnCNN)are two excellent algorithms in the field of image denoising.This paper first studies the influence of the activation function in the DnCNN algorithm,and proves that replacing the ReLU activation function with the PReLU activation function makes the DnCNN network have faster convergence speed,lower loss function loss value and higher peak signal-to-noise ratio(Peak Signal).-toNoise Ratio,PSNR).Secondly,using the prior knowledge of non-local self-similarity of BM3 D image denoising algorithm and the improved feature extraction ability of DnCNN neural network,a block-based deep convolutional neural network image denoising algorithm(BMCNN)is proposed..Firstly,the parameter setting and denoising performance of the denoising algorithm are compared and analyzed by experiments.The BM3 D image denoising algorithm is used to pre-process the noise image.Secondly,the denoised image is used as the estimated image and the original.The noise image is blockmatched separately and used as the input of the DnCNN network.Finally,the image denoising model is established by using the convolutional neural network.The experimental results show that compared with other classical excellent denoising algorithms,the BMCNN image denoising algorithm proposed in this paper shows better denoising performance in the objective evaluation index PSNR.In particular,BMCNN can not only denoise images with repetitive structures,but also have better denoising effects in image denoising of irregular structures.