Single Image Super-resolution Based On Convolutional Neural Networks

In the field of computer vision,image super-resolution reconstruction has been a research hotspot,it has wide application prospect and application value.Its main concern is that in the case of limited performance of the hardware device,the reconstruction algorithm designed from the single or multiple images obtained must not only improve the resolution of the image but also maximally recover the detail of the image's detail features to meet the user's visual requirements for the image.Image super-resolution reconstruction algorithm based on learning is the most popular super-resolution reconstruction algorithm,this kind of algorithm by training images in the library to get the non-linear mapping between high-resolution and low-resolution images to predict a lot of high-frequency information lost in the low-resolution image.Currently,the learning-based super-resolution reconstruction method still has a lot of room for improvement.Deep learning is an emerging machine learning algorithm.Compared with other machine learning methods,its advantages are as follows: 1)Strong ability of processing data.The recognition results in words and sounds show that method based on the deep learning can greatly improve text and voice recognition performance;2)The ability of handling large-scale training data.It is an effective tool to deal with big data.Inspired by this,the paper introduces deep learning algorithm into the problem of super-resolution reconstruction,and aims to improve the quality of super-resolution image by designing a algorithm based on deep learning.The main content and innovation points of this paper are as follows:1.Introduced the relationship between image gradient and image texture in detail,and the relationship between convolution kernel and gradient operator of convolutional neural network is introduced too.Then described the essence of convolution neural network for image feature extraction.And in theory,it shows the superiority of convolutional neural network in super-resolution image reconstruction.2.Propose a four-layer convolutional network reconstruction model with multiple feature maps input,and through continuously changing the network parameters and training network model,the experimental results show that the model has a good ability for reconstruction of single-frame images,and can recovery image's edge and the details of the characteristics well.For a super-resolution reconstruction network model,the ability to extract the detail features of the image is directly related to the quality of the entire reconstruction network,and the extraction of image detail features by the network is closely related to the setting of the parameters of the network layer.Therefore,In the process of network training,obtained a satisfactory set of parameters by continuously adjusting the parameters of convolutional layers in the network finally.Experimental results show that this model can achieve good results in image reconstruction.3.In order to further improve the ability of image super-resolution reconstruction,based on the characteristics of multiple map input the four layers of network reconstruction model,by combining the onvolutional layer's output in front of reconstruction layer used as reconstruction layer's input to improve the reconstruction layer,this kind of input maximizes the characteristics of the image and facilitates the reconstruction of the image.4.In order to study the effect of convolutional neural network on the reconstruction of image with simple features,we obanted two models,model one is trained by only using face image and model two is trained by using other complex image.The two models are used to reconstruct different images respectively and the experiment results show that the network trained by face images is better for face reconstruction,but the generalization of this model is poor.