POCs Like Deep Convolutional Networks For Single Image Super-resolution

Recently,convolutional neural networks have been widely adopted for the single image super-resolution task,and achieve significant improvements over the traditional super-resolution methods in terms of both objective quality and subjective quality.As a classical inverse problem,the critical part for super-resolution is to design more efficient prior models to estimate the potential high resolution structures from the low resolution images.With the rapid development of deep learning techniques,the super-resolution methods based deep learning can be primarily divided into three stages: the shallow networks,deep networks and attention-based networks.At the first stage,the shallow networks are adopted for generalizing the traditional learning-based super-resolution methods,but existing a gradient vanish problem.The deep networks at the second stage employ various approaches for solving the gradient vanish problem and thus build very deep networks.However,the networks at this stage ignore the importance of different regions in the feature map,so the representational ability of the network is powerless.In the third stage,the attention-based networks can adaptively select the interesting regions in the feature map and thus enhance the representational ability of the network.For the high-level vision tasks,adopting the attention mechanism can emphasise discriminative features to improve the recognition accuracy since the importance of each location with feature maps is different.However,it may be suboptimal if applying the attention mechanism in low-level vision tasks(e.g.,image superresolution)because each location with feature maps is equally important.In this sense,it is necessary to design the corresponding feature recalibration module for the super-resolution tasks.To resolve these problems,we visualize the feature maps between the attention-free and attention-based super-resolution networks.An interesting phenomenon can be observed that the networks which extract less noise feature maps trend to achieve the superior superresolution performance,which indicates that the attention module has a certain feature denoising effect,compared to the attention-free super-resolution models.This may be because the attention mechanism extracts the principle component(high energy)of the feature map,which represents the low-frequency information to some extent.In this sense,the attention module has an effect on emphasising low-frequency features and suppressing highfrequency features.However,it loses some subtle textures to some extent when suppressing high-frequency information.This inspires us to design the specific feature denoising module for the super-resolution problem.Since the convolution operation can be regarded as the transform from the view of typical signal processing,suppressing the noise of feature maps is actually denoising in the transform domain.In this sense,we propose two novel lightweight feature denoising blocks: the soft thresholding module and the adaptive soft thresholding module,which generalize the soft thresholding technique,the classical method that denoising in transform domain.By introducing the feature denoising block into the classical attention-free super-resolution network EDSR,we propose a novel projection onto convex sets(POCs)like deep convolutional super-resolution(PL-DSR)network.Experimental results demonstrate that PL-DSR contains more informative features on the basis of low level noise.Therefore,the method proposed by this thesis not only has less model parameters,but also achieves the state-of-the-art performance on the super-resolution tasks.In addition,we expand the proposed method to the field of computational spectral imaging(28.8% of EDSR,74.6% of RCAN and 76.0% of SAN),and obtain better spectral image reconstruction quality than the existing methods.