| Due to external interference,the quality of the image is often degraded during the process of generation and transmission,and the high resolution cannot be maintained.High-resolution images play a very important role in various fields,such as medical treatment,surveillance and so on.With the development of deep learning,convolutional neural networks have made major breakthroughs in the field of image super-resolution.This paper focuses on the single image super-resolution algorithm based on convolutional neural network,and tries to solve the shortcomings of the existing super-resolution network and improve the performance of the algorithm.The work of this paper can provide new suggestions for the network design of single image super-resolution,on the other hand,it also provides new research ideas to complete the task of super-resolution.The paper carries out the following work:Firstly,the degradation model of low resolution image is introduced ? Then several network structures are introduced,including linear structure,residual structure,dense connection structure,multi-path structure,feedback structure and Gan structure? Then it introduces several loss functions?Finally,the attention mechanism widely used at present is introduced?Secondly,SRCNN,FSRCNN,ESPCN,VDSR,SRGAN,EDSR and MSRN are introduced and the advantages and disadvantages of these algorithms are briefly analyzed.Then it summarizes several up-sampling methods,and finally introduces the evaluation criteria,including subjective evaluation and objective evaluation.Thirdly,simple feedforward structure will lead to partial loss of information flow and can not obtain useful features.The use of fixed size convolution kernel will lead to fixed receptive field and can not improve the ability to obtain features.In order to solve the above problems,a singleimage super-resolution network based on feedback mechanism is proposed,which uses an improved receptive field module in the process of feature extraction to expand the receptive field.The feedback mechanism is used to save the memory information of the network,and the attention mechanism is used to redistribute the features.Experiments show that the proposed algorithm can achieve the good results and maintain a balance in the amount of parameters and computation.Finally,as the number of network layers deepens,it will make the model difficult to train.Highfrequency information and low-frequency information are treated equally,which makes the representation ability of neural network insufficient.In order to solve the above problems,an attention mechanism-based single image super-resolution network is proposed.A multi-attention block is introduced to give different weights to different features.The multi-attention block introduces channel attention and spatial attention,so that the network can adaptively focus on features.Experiments show that the proposed algorithm can achieve good results in multiple test sets. |
