Research On Restoration Of Multiple Degraded Images

In the process of image acquisition,the image quality is reduced by various factors,which is called image degradation;If there are multiple types of degradation,such as noise,distortion,defocus,jitter,etc.,this is called multiple degradation.Image restoration is a technique to recover the original image from the degraded image based on the prior knowledge of the degradation process,which has been widely used in satellite remote sensing,public security,biomedical and other fields.In recent years,deep learning methods have been outstanding in image restoration tasks,but most of them only consider a single type of degradation.Therefore,the emphasis of this paper is to use deep learning to solve different types of multiple degraded image restoration.The main work of this paper is summarized as follows:(1)In response to the multiple degradation restoration requirements such as blur and low resolution,a cascade structure network is used.Here we focus on designing an end-to-end multi-scale encoder-decoder network(MED-Net).First,a two-bypass residual block(TRB)is proposed as an important component module of the proposed codec.The two branches of the module use different convolutional layers to detect image features at different scales,and the two features are spliced so that the information of the two branches is shared by the latter layers;Second,add skip connections between the encoding and decoding modules to prevent the gradient from disappearing and accelerate the flow of information;At last,the restoration results are optimized with the detail learning method.Through experimental verification on the public test sets,MED-Net is more effective for non-uniform blurred image restoration tasks that traditional algorithms are difficult to solve;Compared with the advanced joint restoration blur and low resolution multiple degradation algorithm proposed in recent years,the cascade network in this paper has been improved in different evaluation indexes(2)A multi-level attention network(MLA-Net)is proposed to solve the image degradation and restoration problem caused by multiple unknown degradation factors in most scenes.First,the residual branch and information distillation branch were designed under the inspiration of predecessors to extract the high-dimension and low-dimension features of the image respectively;Then,the gate module is used in MLA-Net to adaptively fuse multi-level features;Finally,an improved multi-channel attention selection module(MASM)is added to the output of the network to allow the network to learn more features related to clear results.In the experiment,images with three degradation factors including noise,blur,and compression artifacts were tested.On the test set with different degrees of degradation,the peak signal-to-noise ratio and structural similarity of the MLA-Net restoration results are better than other advanced algorithms;In visual sense,the texture of the MLA-Net restored images is clearer and closer to the original images;In addition,this paper also selects some degraded images outside the data set for testing,and also obtains good restoration results.In summary,taking MED-Net as a key part of the cascade network is very effective to restore non-uniform blur and low-resolution multiple degradation;The single network MLA-Net in this paper can better deal with multiple degradations caused by multiple combined factors such as noise,blur,and compression artifacts.