Model Inspired Learning Methods For Image Denoising

Image denoising is a fundamental problem in image processing and plays an important role in many high-level vision tasks.The purpose of image denoising is to recover the original image from the corrupted image,which is a widely studied but largely unsolved problem.Nowadays,most image denoising methods are learning based,which have achieved better results than the traditional methods and attracted much research attention.But these methods are imperfect,because they are data-driven approaches and usually rely on the training data.If the training data is not enough,the learned function may be over-fitted.If the training data is biased,the decision boundary of the learned function may deviate from the actual distribution.Although traditional methods are getting less popular,the solving process of their optimization models provides good priors to protect the image structure when removing noise.Therefore,this dissertation aims at mining these priors and applies them to the designing of learnable denoising process,which adds some regularization constraints to the denoising process.Such denoising methods can better avoid over-fitting and overcome the impact of data bias by reducing the dependence on training data to some extents.Motivated by the loop unrolling in solving optimization models of traditional denoising methods,we have carried out in-depth research works combining the advantages of data-driven approaches.A model inspired ensemble learning methods for image denoising has been proposed:(1)A flexible ensemble framework for image denoising has been proposed by combining ensemble learning techniques,which uses a group of base denoisers to construct a powerful ensemble denoiser.To the best of our knowledge,this is the first paper to study image denoising using ensemble learning technique,which not only shows the potential of ensemble learning for image denoising,but also provides an effective solution with solid performance.(2)Based on different types of image priors,we propose two types of base denoisers based on the wavelet shrinkage method with shrinkage function constructed by summation of Gabor basis functions,which provides some basic ideas for designing other base denoisers.A model inspired deep learning methods for image denoising has been proposed:(1)An image denoising network with interpretable components inspired by traditional methods' optimization process is proposed to utilize the strong learning ability of neural network.This research work can be viewed as a continuation and extension of the previous ensemble denoiser's study towards the direction of deep learning.The proposed network not only has good interpretability but also achieves the state-of-the-art performance on public datasets.(2)Considering the relation between 1D filter and 2D filter,a novelty directional convolution scheme is introduced in the proposed network.Embedding directional convolution into network can not only reduce lots of parameters,but also provide us another effective way to protect image edges and textures without incorporating time-consuming nonlocal operation.We tested the proposed methods on several popular public image denoising datasets and compared with other state-of-the-art methods.The experimental results verified the effectiveness of our proposed methods.The research works of this dissertation not only show great significance to the development of image denoising methods,but also widen the mentality of their theories and algorithms.