Research On Image Super-resolution Reconstruction Based On Deep Residual Learning

Image spatial resolution refers to the capability of the sensor to observe or measure the smallest object,which depends upon the pixel size.As two-dimensional signal records,digital images with a higher resolution are always desirable in medical,remote sensing,military and other fields.However,in practical applications,resolution is often sacrificed to some extent in order to ensure long-term stable operation of imaging equipment and provide appropriate frame rate for dynamic scenes.Therefore,the improvement of resolution is very necessary.Limited by the technical level and cost of current imaging equipment,people have to improve the image resolution via super-resolution technology.In recent years,the superresolution reconstruction algorithm based on residual learning has made great breakthrough and become one of the most popular image reconstruction algorithm.However,the existing residual structure introduces too many parameters,which makes it difficult to apply in practice.To solve the problem mentioned above of existing residual structure,this thesis has proposed a single image super-resolution reconstruction algorithm based on multi-branch residual structure.The multi-branch architecture is introduced in this novel algorithm to improve the problem on the number of parameter in existing residual structure.Then,according to the characteristics of multi-branch architecture,the residuals module is redesigned in this thesis.The two-layer convolution layer residuals block commonly used in super-resolution reconstruction algorithm is replaced by three-layer residuals block.In order to ensure the image reconstruction quality of the algorithm,this thesis introduces the optimization idea of Enhanced Deep Super-Resolution Network(EDSR)algorithm,and removes Batch Normalization(BN)layer.At the same time,the constant scaling layer is removed according to the structural characteristics of the algorithm.Finally,two reconstruction models based on multi-branch residual structure are designed and implemented.In order to test the performance of the reconstruction model based on multi-branch residual structure,a comparative experiment is carried out on Set5,Set14,BSD100,DIV2 K and other high definition datasets.The experimental results show that in a case of the same quality of image reconstruction,the model designed in this thesis has a smaller number,faster training speed and running speed.This phenomenon shows that the proposed algorithm can effectively solve the problems of existing residual structure and reduce the demand for hardware resources,which indicates that this novel super-resolution algorithm has potential to practical application.