Research On Image Super Resolution Algorithm Based On Convolutional Neural Network

High-resolution images have a wealth of detailed information that can be of great help in image processing and computer vision tasks such as video surveillance,medical imaging,and satellite remote sensing.However,due to limitations of imaging hardware devices,low-resolution images can only be obtained in many cases,and high-resolution image resources are relatively scarce.Therefore,in order to get rid of the limitations of hardware conditions,it is a good choice to use image super-resolution reconstruction technology to obtain high-resolution images.At present,in the field of super-resolution reconstruction,the learning-based algorithm achieves better reconstruction effects.However,most of these algorithms take a long time to train the model,which results in the efficiency of the algorithm is not very high.With the rise of the deep learning boom,convolutional neural networks have been successfully applied to image super-resolution reconstruction.Due to its simple feedforward network structure,the efficiency of super-resolution has been greatly improved.Therefore,this thesis combines convolutional neural networks to study super-resolution algorithms.The main work is as follows:Firstly,an image super-resolution algorithm based on convolutional neural network is introduced.The algorithm uses a convolutional neural network to construct an end-to-end mapping relationship between low-resolution images and high-resolution images,and uses this mapping to complete the reconstruction.Through experimental comparison,the algo-rithm has a good reconstruction effect.Secondly,aiming at the real-time problem of super-resolution reconstruction based on convolutional neural network,an improved image super-resolution algorithm based on con-volutional neural network is proposed.The improved algorithm does not require pre-processing,inputs the original low-resolution image into the network,and performs upsampling process-ing at the network output.At the same time,the optimization algorithm is changed,and the adaptive moment estimation optimization algorithm replaces the stochastic gradient descent optimization algorithm.In addition,a new activation function is constructed to improve the quality of reconstruction.Through experimental comparison,the improved algorithm not only shortens the training time,but also achieves better super-resolution effects.Finally,in order to further improve the effect of super-resolution reconstruction,this thesis deepens the network of small-scale super-resolution algorithm based on convolution-al neural network,and focuses on the experimental study of feature mapping structure,and proposes an image super-resolution algorithm based on deep convolution feature learning.Through experimental comparison,the reconstruction quality of the algorithm is not only further improved,but also improved in real time compared with other algorithms.