Research On Image Super-resolution Algorithm Based On Convolutional Neural Networks

Image super-resolution reconstruction is the process of recovering a high-resolution image from a low-resolution image and is an essential image processing technique in computer vision.Besides simply improving the image perception quality,it helps to enhance other computer vision tasks.It is widely used in many image-related fields such as public safety,remote sensing satellite images,medical images,video display,etc.In general,image super-resolution reconstruction tasks are very challenging and ill-posed.Although the process from HR images to LR images is unique if a degradation model is determined,the mapping is not unique when recovering HR images from LR images.Each LR image has a large number of corresponding HR images.With the rapid development of deep learning,the deep learning-based image super-resolution reconstruction algorithm has been developed significantly.In this dissertation,based on the previous work,the image super-resolution reconstruction based on the convolutional neural network from the aspects of attention mechanism,codec structure,neural network search and contrast learning is studied,and a variety of image super-resolution algorithms based on the convolutional neural network are proposed.The core research work of the dissertation mainly includes the following four parts.(1)An enhanced dual-path attention network is proposed for the contemporaneous convolutional network structure that cannot fully extract the hierarchical features possessed by low-resolution images.The algorithm balances the advantages of residual networks and densely connected networks with a channel attention mechanism.The algorithm utilizes the dual-path network module to fuse the benefits of the residual and densely connected network modules and further enhances the representational capability of the dual-path network through the early separation of features.The network adaptively refines the extracted features by stacking such enhanced dual-path modules and introducing the attention mechanism into the enhanced dual-path modules.Experimental results show that the algorithm has stronger representational capabilities and can reconstruct better visual outcomes.(2)For the task of real-world image super-resolution reconstruction,a robust and effective algorithm is proposed for real-world image super-resolution.This algorithm designs an encoder-decoder residual network structure that progressively reduces noise and recovers lost information.The algorithm further encodes the original image into features with more contextual information by employing an encoder-decoder structure to capture the relationships among an extensive range of pixels.For the encoded features,the algorithm introduces the coarse-to-fine idea into the network to gradually recover high-quality images due to the different spatial scales of the features.The proposed structure can model the lost information and the unnecessary noise at each scale by residual learning.In addition,this algorithm demonstrates that applying batch normalization only to the downsampling and upsampling convolutional layers can lead to performance gains.Finally,qualitative and quantitative experiments show that the algorithm is robust enough to be applied to image super-resolution tasks with unknown degradation processes.(3)A plug-and-play single-image super-resolution algorithm based on neural network search is proposed to address the heavy time-consuming of manually designing neural networks.This algorithm can search network structures with a single RTX 2080 TI GPU.This algorithm searches not only the operation of each node but also the activation function,from-node and skip-connected node,thus exploring a more diverse network structure with less search consumption.This search space implicitly optimizes the number of intermediate nodes and directly avoids the skip-connection explosion phenomenon in other neural network search algorithms.In addition,to eliminate the effect of inconsistent objective functions in the training and testing process,this algorithm introduces random variables to the architectural parameters as regularization during the training process.The searched network structures show that our algorithm can explore more diverse network structures.Qualitative and quantitative experiments on the searched network structures show that the algorithm can search for effective and robust network structures.(4)A lightweight involution-based comparative learning network is proposed for the problem that simultaneous image super-resolution algorithms ignore the spatial location differences of input features.This algorithm firstly applies involution to SISR to introduce spatially specific features.However,the features extracted by the original involution are spatial-specific but channel-agnostic.Therefore,the algorithm further improves the model’s structure to address the original Involution problems.Also,the algorithm optimizes the whole network by mixing L1 loss with self-supervised contrastive loss.The qualitative and quantitative results show that the algorithm achieves state-of-the-art performance.This performance demonstrates the effectiveness and robustness of the algorithm.