| Single-image super-resolution(SISR)based on convolutional neural networks has made great progress in recent years.However,existing methods have started to increase the depth and width of the network to achieve better performance,which has led to excessive computational load and thus difficult to be applied in mobile devices and the real-world applications.This thesis mainly focuses on how to make full use of the extracted low-resolution image information to accurately recover the missing details and reconstruct the high-resolution image with a small number of model parameters and computational constraints.In a word,three main algorithms are designed:(1)Previous methods tend to ignore that the activation function is prone to the loss of intermediate features,thus it is a great challenge to make full use of intermediate features under the constraints of limited parameters and computational volume.In view of this,this thesis proposes a lightweight and efficient feature distillation interaction weighting network(FDIWN).Specifically,by introducing a wide residual mechanism,the wide identical residual weighting unit and a wide convolutional residual weighting unit are used as the basic units for feature extraction.And an efficient wide residual distillation connection framework and self-calibration fusion unit are used to interact features at different scales more flexibly and efficiently.Experiments on five benchmark test datasets,Set5,Set14,BSDS100,Urban100 and Manga109,have fully validated the effectiveness of the algorithm.(2)Transformer can model global dependencies in images through its global feature extraction capability,which is beneficial for the image super-resolution task.Therefore,a feature interaction weighted hybrid network(FIWHN)is proposed in this thesis to reduce the impact of intermediate layer feature loss on the reconstruction quality under the limitation of model lightweighting index.To improve the performance of image super-resolution by combining CNN and Transformer rationally.Specifically,the CNN part still mitigates the problem of intermediate layer feature loss through a wide residual mechanism within the first algorithm.To complement the loss of global features in the CNN model,a new approach to fusing CNN and efficient Transformer is subsequently explored.Experiments on five benchmark test datasets and on selected downstream tasks have fully validated the effectiveness of the algorithm.(3)Given a fact that current Transformer-based approaches ignore the need for the Transformer to combine contextual information for extracting features dynamically,this thesis proposes a lightweight cross-recetived focused inference network(CFIN).The network consists of a series of CT blocks with a mixture of CNN and Transformer.Within each CT block,the model is first enabled to better focus on potentially useful information through a CNN-based cross-scale information aggregation module to improve the feature extraction efficiency in the Transformer phase.The cross-recetived field then guides the Transformer by using a modulated convolutional kernel that understands the current semantic information and exploits the interaction of information within different self-attentions to enable the selection of contextual information required for image reconstruction.Experiments on five benchmark test datasets and two real-world datasets have validated the effectiveness of the algorithm. |
