| Image super-resolution aims to reconstruct high-resolution images from low-resolution images,which has attracted much attention as an important direction in the field of image restoration.The existing super-resolution models have better reconstruction effects,but they often have a large number of parameters and calculations.To solve this problem,this thesis uses knowledge distillation to deeply mine the learning ability of lightweight models and improves the performance of lightweight models through knowledge extraction and migration.This thesis proposes two lightweight super-resolution models and builds an image superresolution system,the main research contents are as follows:Firstly,this thesis proposes an image super-resolution method based on deep feature correlation.To reduce the amount of the calculation of the convolution operation,this method utilizes an improved inverted residual module to replace the standard convolution layer.At the same time,a knowledge distillation framework is built,which contains a converged teacher model and an unconverged student model,and a deep feature correlation module is proposed to mine the deep feature significant relevance knowledge of the teacher model,and then transfer it to the student model for learning.The experimental results show that the knowledge distillation framework proposed by this method can help the lightweight student model improve the quality of super-resolution image without changing the network structure.Secondly,this thesis also proposes an image super-resolution method based on multichannel distillation.This method also replaces the standard convolution layer with the linear feature transform module.To solve the problem of information loss caused by the linear feature transform module,a multi-channel distillation module is introduced to retain more deep feature information.At the same time,the attention mechanism is introduced to highlight important feature,helping the reconstruction module to focus on the restoration of highfrequency details.Experiments show that the method proposed in this thesis can greatly reduce the amount of parameters and calculations of the model,while effectively improving the super-resolution effect of the model.Thirdly,this thesis designs an image super-resolution system,and it contains multiple super-resolution methods.The system interface is simple and easy to operate,which is convenient for users to understand and use the image super-resolution algorithms. |
