Research On Image Super-resolution Algorithm Based On Deep Learning

As one of the most common forms of information,image is widely used in various fields.The resolution of an image is an important parameter that reflects the amount of information.However,in many practical applications,the resolution of an image is usually hard to meet the demand,so the image super-resolution algorithm comes into being.The target of image super-resolution algorithm is to efficiently reconstruct the corresponding high-resolution images from low-resolution images.Due to the advantages of high flexibility and low implementation cost,image super-resolution algorithm is gradually becoming a popular research direction.With the development of deep learning,image super-resolution algorithms based on deep learning are showing strong reconstruction performance.Nowadays almost all the image super-resolution algorithms based on deep learning build convolutional neural networks for end-to-end super-resolution processes.Therefore,the architecture of convolutional neural network largely determines the performance of the algorithm.This thesis explores the design of convolutional neural networks for image super-resolution tasks,and aims to the research of image super-resolution networks with high reconstruction performance and high reconstruction efficiency.The main contributions are as follows:Firstly,in order to learn and make full use of adequate hierarchical features for recovering more detailed information,this thesis proposes a fast and accurate deep fractal residual network.The network uses fractal expansion rule to establish fractal block which contains multiple different feature extraction paths,so that different hierarchical features can be extracted and then fully utilized through the fusion operations.Thus,the fractal blocks learn and combine different hierarchical features to generate finer features favoring the reconstruction of high-resolution images.Moreover,bothlocal residual learning and global residual learning are applied to preserve the low-level features and decrease the difficulty of training.Finally,this thesis proposes a weightsharing model with fewer parameters to reduce the space complexity while keeping comparable performance.Experimental results demonstrate that the network proposed in this thesis can effectively extract various hierarchical features and show good image super-resolution performance on public datasets.Secondly,in order to further reduce the model size and improve the super-resolution performance of network,this thesis proposes a lightweight image super-resolution network from the perspective of receptive field and frequency information.The network uses de-subpixel operation to change the spatial size of feature maps,which not only improves the network's ability to extract features of different frequency information,but also significantly reduces the network's computational complexity,thus taking into account both reconstruction performance and model size.In order to further reduce the model size and improve the practicality,this thesis builds and trains the network in a multi-scale manner,so that a single model can be applied to the reconstruction processes of different scaling factors.The network improves efficiency while maintaining competitive performance in full-reference quality metrics and visual quality,and achieves a better balance between reconstruction performance and model size.Thirdly,this thesis proposes a lightweight attention mechanism and designs an image super-resolution network named residual dense attention network based on this attention mechanism.The attention mechanism combines channel-wise attention and global/local spatial attention,which can focus on different channels and different spatial positions of the feature maps,thereby helping to highlight the important information for the super-resolution process.In addition,both residual learning and dense connections are introduced into the network to further preserve the features extracted by the shallow layers and achieve feature reuse.The proposed attention mechanism can significantly improve the network's performance at the cost of a small extra amount of operations and parameters.Moreover,the network proposed in this thesis has better reconstruction performance than networks with similar model sizes,and can deal with more complex super-resolution tasks and recover more accurate image details.