| In the process of image acquisition and transmission,it is inevitable to be polluted by various factors,which cause image degradation and loss of detail.This not only seriously reduces the quality of the image,but also directly affects the subsequent image processing tasks which require high quality of the image.Therefore,restoring image from these degraded images is an important and long-standing problem in digital image processing.In recent years,the recovery of degraded images based on deep learning has become a hot research topic and achieved more advanced results than traditional methods.Therefore,this paper conducts an in-depth study on JPEG compression artifacts reduction and image denoising based on deep learning.Specific tasks are as follows:Firstly,targeting the shortcomings of image degradation caused by the JPEG compression process,a jointly image filter containing two lightweight sub-networks is proposed.To effectively transfer the common structure of the input and guided image,the first sub-network uses a reversible down-sampling layer to enlarge the receptive field and extract feature information to construct a guided feature image.Then the input image is further fused with guided feature image as the input of the second sub-network.Finally,the common structure is mapped to the recovered image based on residual learning.The network jointly trains images at various levels of compression,overcoming the limitations in the application of fixed models.The proposed algorithm not only effectively removes the JPEG compression artifacts,but also extracts clearer edge structures.Secondly,due to the problem of poor flexibility and practicability of the specific denoising model,a denoising network based on a dense connection module and residual learning is proposed.This method performs dense connections on dilated convolution to construct dense module,densely extracts image features while enlarging the receptive field.The representation capability of the network is improved by extracting multi-scale and multi-level features,while training a blind denoising model containing a large range of random noise using a noise level matrix.This design effectively improves the overall generalization capability of the network.The experimental results show that the proposed model preserves the structural details well with better visualization than several advanced algorithms.Finally,to address the problem that traditional guided image filters are difficult and inaccurate to solve for the coefficients,a novel network learning method which predicts the linear representation coefficients based on convolutional neural network is proposed.To transfer structural features more effectively,the method does not directly predict image,but rather constructs residual dense block and multi-scale structures for predicting the linear representation coefficients,which are used to indirectly represent the residual image.Such design effectively remove noise while recover fine structures and edges better.The proposed method could deal with different noise levels and spatially variable noise and has a practical application value. |
