Research On Deep Learning Based Single Image Super Resolution

As an important area of signal processing,image super-resolution reconstruction has been widely used in medical image processing,satellite remote sensing image processing,high-definition digital TV,video monitoring and other high-tech fields,which has great industrial application and academic research value.In recent years,with the great progress of convolutional neural network in the field of image recognition,deep learning is gradually introduced into single image superresolution task.However,most deep models treated different types of information equally(e.g.,spatial position or channel independent),and perform badly when recovering high frequency details,resulting in over-smooth and lacking of textural information in the output.On the other hand,as the depth of network grows,information in shallow layers is not thoroughly utilized for feature mapping,resulting in the performance by simply stacking convolution layers can not be significantly increased,and the extremely deep network is difficult to train.Based on the above problems,this paper proposes a novel up-sampling method based on residual inverse discrete wavelet transform and a loss function that supervises both the spatial and the frequency domain.Combined with attention mechanism and twolevel folding structure,the wavelet channel and spatial attention network finally proposed in this paper achieved the state of the art performance in several commonly used test sets.Specifically,this paper has the following contributions:(1)This paper constructs a robust single image super-resolution baseline model,which lays a solid foundation for the follow-up research.The research of baseline model is mainly divided into three sub works.In this paper,two new up-sampling methods are proposed.The proposed residual inverse discrete wavelet transform can learn different high frequency and low frequency sub-bands respectively,and the high frequency components that are easy to be lost are easier to be captured due to the learning of high frequency sub-band.In this paper,a new loss function is proposed to supervise the loss function of both spatial and frequency domain,which makes the supervision of different frequencies more balanced,which makes up for the weakness of the loss function in the past which only supervises the spatial domain to supervise more high-frequency information;we also verify that the batch normalization layer is not suitable for image super-resolution task,and improves the learning rate by introducing the weight normalization layer into the base block,which greatly accelerates the training speed of the network.Through the above new proposed technology,the baseline model proposed in this paper achieves competitive results compared with the state of the art models before 2018;(2)This paper studies the method of integrating attention mechanism into image super-resolution task.By distinguishing different features with different attention,the hard features can be learned better.Specifically,the improved channel attention layer greatly simplifies the transmission of information flow,making the highfrequency channel easier to be transmitted.The spatial attention wide activation layer proposed in this paper can make different channels get better frequency focus,that is,the high-frequency features in the low-frequency filter can be killed by the activation layer(the same with the high-frequency filter),so that both the highfrequency filters and the low-frequency filters can get better learning.After the two attention mechanisms are integrated into the basic block,the performance parameter trade off of the baseline model is greatly improved,and the proposed model reaches the state of the art performance in the existing lightweight image super-resolution network;(3)The hierarchical sequential fusion module proposed in this paper enables the features of different depths to be captured in the later stage.The shallower features pass through more channel compression layers,and the more valuable deep features are better preserved.The network depth can be easily increased through the secondary folding structure,and the shallow features and deep features hided in the whole network pipeline can get better mining.After effectively stacking a large number of base blocks,the final model proposed in this paper achieves the state of the art performance on several super-resolution common Benchmarks.