Research On Noise Intensity Estimable Image Denoising Algorithm Based On Deep Learning

Images are carriers of information and are an important source of information.However,due to the presence of noise,images become blurred and important details are lost,which is not conducive to accurate information acquisition and affects people’s visual perception,so the research on image noise reduction algorithms is of great significance.With the development of artificial intelligence and the increasing power of GPUs,deep learning based image noise reduction algorithms have been proposed and have achieved excellent results.However,there are still many problems,such as the fact that most of the current algorithmic models deal with noisy images of a single noise intensity,and a model trained with a single noise intensity cannot deal with images of different noise intensities at the same time.In addition,although deep learning is used for image noise reduction,the objective indexes are continuously improved,but the texture details of the processed images are too smooth and not well recovered.In order to effectively solve the above problems,this paper is based on deep learning to study the above problems,the research content is as follows:(1)In order to make the model capable of processing multiple noise images of different intensities simultaneously and improve the robustness of the noise reduction task.This paper designs a noise intensity classification model based on convolutional neural networks.The model is able to accurately identify seven different noise intensity images.After the classification model has estimated the noise intensity,the corresponding noise reduction algorithm is selected to carry out noise reduction,so that the model can carry out noise reduction for a variety of noise images of different intensities at the same time.For the problem of poor interpretation of deep learning,this paper performs feature map visualisation and generates a heat map using the Grad-CAM algorithm.(2)For most deep learning models after noise reduction of images,there are problems such as loss of texture details.In this paper,in order to improve such problems effectively,the DNCNN noise reduction algorithm is improved and a multi-scale densely connected image noise reduction algorithm is proposed.For the extraction of shallow features of the network,inspired by the Inception connection structure,this paper adopts a multi-scale feature fusion technique,which expands the perceptual field of the network and reduces the loss of information.To solve problems such as degradation of the model as the depth of the network continues to deepen,inspired by the Dense Net connectivity,the connectivity structure of layers 2 to 15 of the DNCNN is improved by introducing a dense connectivity structure with a total of 15 densely connected units,making each layer connected to all the previous layers and enhancing the information transfer.In addition this chapter carries out a visualisation analysis by plotting the residual image obtained from the noisy image after the noise reduction model,the maximum response image of the filter obtained using the gradient ascent algorithm,i.e.the convolution kernel visualisation,and the feature map visualisation respectively.(3)After experimental analysis,for the noise intensity recognition model.A comparison of the algorithm in this paper with other established classification architectures shows that the noise intensity classification model designed in this paper has fewer parameters,takes less time to train and runs faster.For the noise reduction model,the proposed multi-scale densely connected image noise reduction algorithm has a higher peak signal-to-noise ratio than other algorithms,and the texture detail information is further preserved.A visual analysis of the noise intensity recognition model and the noise reduction model shows that the model focuses more on the noise point information distribution of the noisy image.