Research On Single Image Super-resolution Based On Attention Learning

With the rapid development of science and technology,images,as an important carrier of information transmission,have been integrated into people’s daily lives and play an important role.However,due to the limitations of imaging devices and external environment,the acquired images often have different levels of blur and noise,which seriously affect the visual experience of people and subsequent high-level pattern recognition applications.Although the image resolution can be enhanced by improving the physical device in the imaging system,such kind of method is often limited by cost and working environment.Image super-resolution(SR)is a software-based method to enhance image resolution,which can improve image quality without updating imaging devices.SR has been widely used in remote sensing,medical imaging,entertainment,video surveillance,etc.,and has attracted great attention from scholars all over the world.This thesis focuses on the learning-based single image SR reconstruction,which can learn the mapping function from a large amount of external LR-HR training image pairs.At the same time,new studies in SR reconstruction have also been introduced,such as high-resolution image information,attention mechanism.The main contributions of this paper are as follows.(1)A novel high resolution similarity directed adjusted anchored neighborhood regression method for single image SR is proposed,which effectively reduces the impact of non-similarity HR patches on SR performance.The adjusted anchored neighborhood regression(A+)method is one of the state-of-the-art methods for single image SR.An important implicit assumption of the A+ method is that the high-resolution(HR)image patches corresponding to similar low-resolution(LR)image patches must be similar too.Therefore,the neighborhood regressions in HR patch space and LR patch space can share same representing coefficients.However,this assumption is often invalid due to the ill-posedness of the SR problem(i.e.multiple HR images can be degraded to the same LR image),and non-similar HR sample patches often share large representing coefficients.To remedy this,we propose to improve the A+ method by introducing high-resolution similarity-based adjusting weights into HR representation coefficients to reduce the effect of these non-similar HR sample patches.The numerical results demonstrate that our method can improve the performance of the restored HR images effectively with low computational cost.(2)A multi-grained attention network for single image SR is proposed,which fully exploits the properties of multi-scale and dense skip connections,and introduces multi-grained attention mechanism into SR learning.Recently,deep convolutional neural networks(CNN)have drawn great attention in image SR,and achieved an excellent SR performance.However,these methods are usually stacked more convolutional layers.The idea of multi-scale feature fusion has not been effectively exploited,which limits the SR performance.Therefore,we propose a multi-scale dense connection network to efficiently extract the image features at multiple scales and capture the features of different layers through dense skip connections.In addition,we introduce a multi-grained attention mechanism to further enhance the discriminative representation ability of the features.In our method,the importance of each neuron is computed according to its surrounding regions in a multi-grained fashion and then is used to adaptively re-scale the feature responses.More importantly,the “channel attention” and “spatial attention” strategies in previous methods can be essentially considered as two special cases of our method.Experimental results show that our method outperforms state-of-the-art SR methods both quantitatively and qualitatively.Meanwhile,our method can well restore the edge and image details.(3)A pyramidal dense attention network for image SR is proposed,which can effectively explore multi-level features with a few parameters.Due to the inconsistency between the number of input and output channels,the traditional dense skip connections with a fixed channel growth rate can cause the loss of information and degrade the learning ability of the network.To solve this problem,we propose a pyramidal dense connection architecture with a varying growth rate,where the output feature dimensionality can gradually increase as the network layers deepens to extract SR features efficiently.However,this method will result in the parameter explosion.As regard to this,we introduce an adaptive group convolution strategy.Unlike the previous group convolution,the number of groups grows linearly with convolutional layers to further reduce the number of parameters.Besides,we also present a novel joint attention to capture cross-dimension dependencies between the spatial dimension and channel dimension in an efficient way for providing rich discriminative feature representations.The experimental results show that our method can achieve better performance than some other state-of-the-art lightweight SR methods.(4)A novel feedback pyramid attention network for single image SR is proposed.With exploiting feedback mechanism and pyramid non-local structure in an efficient way,the representation ability of the network is improved.Due to the strong learning ability,CNN based image SR methods have achieved significant performance improvement.However,most CNN-based methods mainly focus on feed-forward architecture design and neglect to explore the feedback mechanism,which usually exists in the human visual system.In this paper,we propose a feedback pyramid attention network to fully exploit the mutual dependencies of features in the SR model.Specifically,a novel feedback connection structure is developed to enhance low-level feature expression with high-level information.In our method,the output of each layer in the first stage is also used as the input of the corresponding layer in the next state to refine the previous layer representations.Moreover,we introduce a pyramid non-local structure to model global contextual information in different scales and improve the discriminative representation of the network.Extensive experimental results on various datasets demonstrate the superiority of our method in comparison with state-of-the-art SR methods.The proposed method is robust to different image degradation models.