Research On Image Super-Resolution Based On Convolutional Neural Network

Single image super-resolution is one of the most classic tasks in the field of computer vision.This technology uses the software algorithm to reconstruct the high-resolution image from the low-resolution image,which enriches the details of the image and enhances the visual effect of the image.This technology has broad application prospects in many fields such as public safety,medical imagery and satellite remote sensing.Therefore,image super-resolution has become one of the research hotspots in recent years.With the rapid development of deep learning,image super-resolution algorithms based on deep learning have made great progress.Based on the predecessors' work,this thesis studies the image super-resolution based on the convolutional neural network.The main contributions of this thesis are as follows:1.This thesis briefly summarizes the research background,significance and current research status at home and abroad of image super-resolution.In addition,the image super-resolution algorithms are classified and summarized.After that,the commonly used image super-resolution quality evaluation indicators are given from both subjective and objective aspects,and the relevant theoretical knowledge of convolutional neural networks is briefly described.2.Several image super-resolution algorithms based on the convolutional neural network are discussed emphatically.And this thesis proposes a novel image super-resolution algorithm based on stacked U-shape networks with channel-wise attention.The algorithm enriches multi-scale features by using U-shape network module,and the output of the module is high-resolution feature,which can be directly used to reconstruct an image without using upsampling layers.In fact,few research works study attention mechanism in single image super-resolution and research shows that the attention mechanism can improve the feature representation,the algorithm introduces the channel attention mechanism by modeling the relationship between the feature channels,to suppress the redundant features and further strengthen the useful features.Extensive evaluations on public datasets demonstrate that the proposed method achieves good performance in terms of visual effect and image quality metrics.3.In the task of image super-resolution,the channel-wise attention mechanism can improve the quality of image reconstruction.Compared with the relationship between the feature channels,the spatial position relationship is also very important for image super-resolution task.In view of this,this thesis further introduces the concurrent spatial and channel-wise attention mechanism,and proposes a image super-resolution algorithm based on stacked U-Nets with concurrent spatial and channel-wise attention.The concurrent spatial and channel-wise attention mechanism optimizes the features from both channel and space to improve the reconstruction performance.The U-Net module embedded with the dual attention mechanism is used as the basic module to construct the network in a cascading manner,and the reconstruction performance is further improved.The experimental results show that the proposed method can achieve better accuracy and visual improvements than the state-of-the-art methods.