Face Image Super-resolution Reconstruction Based On Convolutional Neural Network

With the development and progress of the information age,the image has gradually become one of the important ways to obtain and transmit information.The resolution of images also directly affects people’s access to effective information.Therefore,in daily life,people have higher and higher requirements for image quality.They are no longer satisfied with fuzzy images but pursue higher-clarity images.However,in real life,due to the limitations of the quality of image acquisition equipment and environmental factors,only low-resolution images can be obtained,which makes the effective information obtained is also very limited.Therefore,it becomes particularly important to improve the image resolution to obtain more effective information.Usually,software techniques are used to improve the resolution of images,including facial images,a process called image super-resolution reconstruction.The use of the convolutional neural network for image super-resolution reconstruction is the mainstream method in the field of super-resolution reconstruction at present,but there are also many shortcomings.For example,when the depth of the network is too shallow,the feature information extraction ability is insufficient and the field of perception is small,which leads to the blurred texture details of the reconstructed face image.When the depth of the network is too deep,it may lead to difficulties in network training and problems such as gradient disappearing and gradient explosion.Compared to traditional super-resolution reconstruction,facial images have higher similarity and greater difficulty in reconstruction.Therefore,in order to solve these problems,this paper proposes two kinds of face image super-resolution reconstruction algorithms.The first one uses a residual network and introduces the attention mechanism to carry out super-score reconstruction.The second one uses multi-scale convolution instead of single convolution and combines coordinate attention mechanisms for super-resolution reconstruction.The main contents of the face image super-resolution reconstruction algorithm based on the residual attention mechanism are as follows:Firstly,feature extraction by an adaptive residual network.The introduction of residual thought can improve the quality of the network.In this algorithm,an adaptive residual network is added into the network feature extraction module,which is also a residual network in essence.The convolution layer is added to the traditional skip connection path,and a feature supplement is carried out to enhance the richness of output feature information on the basis of realizing residual networks to accelerate network convergence and alleviate gradient problems.Secondly,attention mechanism is used to reduce the interference of redundant information on reconstruction performance.By taking advantage of the feature of focus attention,the attention mechanism can highlight local information,such as mouth,eyes,and other parts of the face image,and suppress the influence of noise,background,and other invalid information on network reconstruction performance.In this algorithm,the convolutional attention module is used to combine channel and spatial attention,so as to generate the attention feature map in both channel and spatial dimensions.Then,the feature map is multiplied with the original input feature map to achieve adaptive feature correction and generate the final feature map.This approach increases the attention to features and improves their reconstruction performance.The main contents of the face image super-resolution reconstruction algorithm based on a multi-scale residual attention mechanism are as follows:Firstly,multi-scale convolution is used to replace single convolution for feature extraction.Multi-scale convolution uses convolution kernels of different scales to extract features of different scale layers of images and then carries out feature fusion.Compared with single convolution,multi-scale convolution not only improves the richness of feature information but also alleviates the loss of image feature information and improves the feature utilization rate of the network.Secondly,coordinate attention is used to improve network attention.This algorithm adds the coordinate attention mechanism after feature fusion by multi-scale convolution extraction.Compared with channel attention and convolution attention module,it pays more attention to location information,so the network can not only obtain cross-channel information but also locate the region of interest more accurately from the direction and position.Improve the network attention to reconstruct image quality better.Finally,the algorithm uses sub-pixel convolution instead of deconvolution to improve reconstruction efficiency.Unlike deconvolution,sub-pixel convolution moves the feature extraction process with a large number of computations to a lower-dimensional space,which can effectively reduce the number of parameters and shorten the network training time.In order to verify the effectiveness of the proposed algorithm,the proposed algorithm is experimentally compared with other convolutional neural network models.The experimental results show that compared with the comparison algorithm,the proposed algorithm has better reconstruction performance and the reconstructed facial images have more details and clearer texture.In the objective evaluation,the proposed algorithm is superior to other algorithms in terms of peak signal to noise ratio and structural similarity.