Research On Image Denoising Algorithm Based On Wavelet Transform And Deep Learning

As a fundamental and important branch in the field of image treatment and computer vision,image denoising is a fundamental and important subfield,it has an extremely vital role in object detection,goal tracking,image feature extraction,etc.Deep learning has brought new research directions for image denoising and has gradually replaced traditional image denoising algorithms by achieving better results than traditional image denoising algorithms.Images are by far still one of the most common ways to acquire information,and with the growth of fields related to the direct acquisition of information through images,higher demands are placed on the quality of images,and therefore on image denoising as well as image denoising algorithms.To build a deep learning-based image denoising network model for the purpose of improving image denoising quality,protecting image detail information,evaluating network architectures and decreasing network parameters,the main work includes the following two aspects:(1)A model that introduces an attention mechanism and dilates the convolution is proposed.In order to improve the image denoising effect in one step,the perceptual field of the whole network is increased by designing the cavity block,so as to obtain more scale information and improve the ability of acquiring image features;T Merging spatial attention mechanism and channel attention mechanism,adjusting in both channel and space to improve feature map;Use Mish function instead of ReLU function as activation function after each convolution layer to increase the capability of overall nonlinear adaptation in the network;the concatenate connection is used to protect the detail information in the original image contained in the noisy image.Experiments with noise levels of 15,25 and 50 are conducted on the image denoising general test sets Setl2 and BSD68,respectively,and are compared and analyzed with other popular denoising methods in terms of subjective evaluation,PSNR,SSIM.The experimental results show that the algorithm can preserve the graphic detail information and enhance the de-noising effect to a certain extent.(2)A lightweight image denoising model based on wavelet multi-component is proposed.In order to reduce the counts of parameters,training time and the number of FLOPs of this network model without degrading the image denoising effect,partial lightweight convolution is adopted.The basic network model adopts ’U’ type network structure,symmetrical jump connection structure,encoder to obtain image feature information,decoder for up sampling,can accurately capture the feature information;the image is decomposed into high and low frequency components by wavelet decomposition.Different numbers of Ghost lightweight convolutions are designed for low frequency components and high frequency components containing most noise and image details.Targeted training protects the border detector information of the image and diminishes the model’s full parameters.BSD68,CBSD,Set14 and Urban 100 are used as test sets to test the model,and compared with other image denoising algorithms in terms of image denoising effect,model parameter quantity and model training time.The results of the experiments reveal that the network model reduces the covariates of the algorithm,training hours and FLOPs of the network model,and the denoising performance is not considerably degraded compared with the predominant denoising methods.