Research On Image Denoising Algorithm Based On Multi-Scale Feature

Image is an important carrier of information transmission,and noise may be introduced in the process of image acquisition,processing,transmission and reception,which hinders information recognition and subsequent image processing.Therefore,image denoising is a classical problem in the field of low-level computer vision.The real noise carried by image is complex in form,diverse in expression and limited in training data.At present,its removal is faced with a series of challenging problems.Among them,with the deepening of network layers,the actual receptive field increases slowly,which leads to the limited information capture range and fails to effectively use the global and neighborhood semantic information to guide the denoising of the target area.Due to the limited training data of images with real noise,it is difficult to make full use of the powerful learning performance of deep neural network.Aiming at above issues,this thesis improves the current mainstream denoising network.The main innovations and contributions are as follows:1.A multi-scale feature based real image denoising network is proposed.There are three innovations in the network structure:(1)Channel attention mechanism is introduced in noise estimation to extract noise features discriminatively,so that the network can strengthen the influence of valuable features and suppress low value or even worthless features adaptively.(2)Use the idea of pooling pyramid to obtain multi-scale content,and denoise at the global and detail levels,so that the image information is fully considered and utilized.(3)For the feature maps with different scale information,the structure of dynamic convolution kernel is used to adaptively select different convolution kernels to express features,and then fuse the features.Experiment results show that the proposed multi-scale feature based real image denoising network,each module of which has a significant contribution to the improvement of the overall network’s denoising effect.2.In order to solve the problem of limited training data of real noise images,(1)This thesis uses limited real noise datasets,adopts an improved data augmentation strategy,and carries out data augmentation without losing content information,so that the network can not only learn "how" to denoise,but also understand "where" to denoise,and denoise different image regions to different degrees to avoid excessive smoothing.(2)By using the generative adversarial network,the united framework of denoiser and noise generator is established to model and simulate the real noise,which further enriches the training data.Experimental results show that using augmented training dataset,the multi-scale feature based real image denoising network has achieved good results in both quantitative and qualitative analysis,and is superior to other mainstream algorithms on the public test dataset SIDD,DND and NC12.