Research On Image Denoising Based On Dilated Convolution And Noise Estimation

With the development of the Internet and the popularity of smart devices,images have been playing an important role in biomedicine,public management and so on.Af-fected by many factors such as environmental influences or defects of electronic imag-ing equipments,images suffer from noise inevitably,which influences the computer vision tasks.Therefore,image denoising is an important subject of study for computer vision.In recent years,the denoising methods based on convolutional neural networks have made notable achievements.However,these algorithms still face many difficul-ties.For example,models which have a large number of parameters are not suitable in practical application;single-stream structures fail to reconstruct fine textures and pat-terns;denoising methods cannot deal with multi-type noises at the same time.This thesis introduces dilated convolution and noise estimation,and studies on deep learning based methods in terms of model size,image quality and denoising ability of multi-type noises.The contents of this thesis are as follows:In response to deep models which are marred by a high number of parameters,we propose a dilated residual encode-decode networks for image denoising.The proposed network uses an end-to-end denoising model,whose key block can extract details ef-fectively using few parameters.Moreover,the residual strategy makes inference and backpropagation more efficient.Extensive experiments on synthetic noisy images are conducted to evaluate the effectiveness of our proposed method,and show that our method leads to remarkable denoising results only using few parameters.To address the problem that a single-stream structure may fail to reconstruct fine textures and patterns,we propose a multi-scale gated fusion network,which makes full use of multi-scale information.We incorporate dilated convolution into a merge-and-run module to exploit multi-scale features in an effective way and further recognize useful features by filtration via a gating mechanism.Moreover,we propose a simple but effective loss function based on the sparse algorithm to boost image visual quality.The extensive experiments on benchmark datasets show that our proposed method can well recover textures.To address the problem that the model trained using a single type of noise cannot deal with other types of noises well,we proposed a denoising method based on the esti-mation of noise type and noise level.The network is a two-stage method.First,we use the noise estimation network to estimate noise type and noise level from noisy images and constructs corresponding masks.Then,we concatenate the masks and the original input to build the mix input that is the input of the two-stream denoising network.The output of the denoising network is a final noise-free image.The extensive experiments on benchmark datasets show the effectiveness of using noise information,and further prove that our proposed method does well in blind denoising of multi-type noises.