Image Super Resolution Reconstruction Based On Attention Mechanism And Residual Feature Fusion

Due to its significant role in production and life,image superresolution reconstruction technology has always been the focus and hotspot of research,which also produces many classical super-resolution reconstruction algorithms.With the development of deep learning and the renewal and iteration of hardware equipment,many methods use the residual structure to build a very deep convolutional neural network to realize image super-resolution reconstruction,but these methods do not make full use of the hierarchical features on the residual branches,resulting in poor information flow between layers in the network,resulting in the inability to further improve the reconstruction effect.In addition,many methods also use the attention mechanism to improve the performance of network reconstruction,but on the one hand,the spatial attention mechanism used in the image super-resolution task is often calculated by convolution.The convolution operation can only capture the local key position information,and can not model the long-term dependence of the image structure in the image super-resolution task.On the other hand,simply building the two attention mechanisms in the network through parallel connection or serial connection will not only improve the image reconstruction quality,but also bring a large number of parameters,resulting in a waste of computing resources.In view of the above problems,the research of this paper is as follows:Firstly,this paper designs a residual feature fusion structure.The whole structure is built by several residual blocks,and the skip connection is used to increase the flow of information within the structure.At the same time,the features on each residual branch are forwarded to the tail of the structure and fused into more representative features,which solves the long-term dependence of high-frequency features and low-frequency features in the whole network.In addition,in the structure,the asymmetric convolution block is used to enhance the convolution kernel skeleton,accelerate the training of the network,strengthen the ability of extracting salient features,and enhance the influence of local salient features on image super-resolution reconstruction.Secondly,this paper proposes a network based on the fusion of coordinate attention and residual features.The whole network uses the residual structure to obtain a very deep trainable network.The residual feature fusion attention block is used as the basic module of the network.The residual feature fusion attention block is composed of enhanced asymmetric convolution block and coordinate attention mechanism.By adaptively adjusting the channel relationship through the coordinate attention mechanism and embedding the spatial information on the original image into the channel attention through the horizontal and vertical directions,it can not only accurately capture the exact position of the important area in the original image,so as to avoid the loss of the longterm dependence in the original image space when using convolution to calculate the spatial attention mechanism,and let the residual features fusion the features extracted by the attention block gather in the important area of the original image.At the same time,it will not bring a large number of parameters and affect the real-time performance of the network.Finally,through various comparative experiments on five public data sets,it is proved that the proposed algorithm has better reconstruction quality than other methods from the subjective and objective perspectives.