Image Super-resolution Using Attention Convolutional Neural Networks

Image super-resolution(SR)is a low-level vision task to recover high-resolution(HR)images from their matching low-resolution(LR)images.In the real world,a large number of images are used to record and save various information.However,due to the influence and limitation of various factors,such as shooting position,ambient light,optical devices,and so on.the acquired image or part of the image may be low resolution or even fuzzy.Therefore,SR is needed to obtain matched high-resolution images close to real images.However,SR is an illposed problem since infinite HR images can be downsampled to the same LR image.Therefore,it is too hard to find a good solution in the extremely large space of the possible functions.Moreover,high-frequency information and texture information which are most needed in SR task are often missing in the LR images.However,LR images contain a lot of low-frequency information to hinder image SR reconstruction.These problems make image SR a challenging task.In recent years,deep learning methods have provided a great boost to computer vision and provided new solutions for image super-resolution tasks.Aiming at the difficulties of the above image super-resolution tasks,this paper proposes two new image super-resolution algorithms based on deep learning algorithms such as convolutional neural networks(CNNs)and attention mechanism.The main work of this paper is as follows:This paper proposes a feedback attention network(FBAN)for SR.In this network,feedback mechanism and dense connections are used to improve the feature extraction capability of the network.In order to make the network focus on high-frequency information and improve the representational ability of the network,a dual attention mechanism is used.And the skip connection of residual learning is adopted to reduce the complexity of network mapping,accelerate the learning speed and convergence speed of the network.Through the combination of multiple mechanisms,FBAN is finally able to achieve better super-resolution results with fewer training iterations and a smaller training set.In addition,it is proved that the network is lightweight and efficient by experimental comparison with various advanced methods.This paper also proposes a residual dense attention network(RDAN)with better performance for SR.In order to maintain the same feature reuse ability as Dense Net,and to alleviate the problem of heavy memory occupation in Dense Net,the residual dense module is used as the main structure of the network.At the same time,a light-weight channel attention mechanism is used to improve the feature extraction and expression ability of the whole network with fewer increasing parameters.By experimental comparison with advanced SR algorithms,RDAN can exceed the super-resolution effect of large networks with fewer parameters and fewer training iterations.It shows that RDAN has excellent performance and high efficiency to implement image SR.