Research On Video Super-resolution Algorithm Based On Multi-attention Feature Fusion

In recent years,with the rapid development of deep learning,super-resolution reconstruction technology has received extensive attention.As the attention mechanism shines in the fields of image segmentation and detection,researchers have gradually introduced the attention mechanism into the super-resolution direction,and have achieved many remarkable results.Based on the attention mechanism,this thesis conducts exploratory and innovative research on super-resolution reconstruction in single image and video,and has achieved the following results.In view of the current super-resolution network,there are problems that channel weight prediction destroys the direct correspondence between each channel and its weight,and the existing networks only consider the first-order or second-order channel attention,failing to fully consider the attention of different-order channels for the impact of network performance.Thus,this thesis proposes an image super-resolution reconstruction algorithm based on a mixed-order channel attention network.First,the network framework uses the local cross-channel interaction strategy to change the dimensionality reduction operation of the previous first-and second-order channel attention model to the one-dimensional convolution implementation of the kernel k.This not only makes the channel attention prediction more direct and accurate,but also simpler than the previous channel attention model.In addition,this model also uses the above improved first and second-order channel attention models to comprehensively utilize the advantages of different orders of channel attention and improves the ability of network discrimination.Compared with the existing super-resolution algorithms on the benchmark data set,the reconstructed image texture details and high-frequency information of the algorithm proposed in this thesis can be better restored,and the perception indicators on the Set5,Set14,BSD100 and Urban100 test sets are improved by 8.2%?6.4%?6.1% and 6.8%.These show that the network adopts more accurate channel attention prediction and comprehensively utilizes different orders of channel attention to have a certain effect on performance improvement.In view of the fact that the attention mechanism has achieved better results in single image super-resolution,this thesis will continue to study multi-frame super-resolution,that is,video super-resolution.In view of the fact that the reference network fails to treat different frame features differently,that is,it fails to pay more attention to the useful features,and the reference network simply concatenates all frame information without making full use of inter-frame temporal relation and intra-frame spatial relation,this thesis proposes a video super-resolution algorithm based on dual attention temporal-spatial fusion(DATS-VSR).Specifically,this thesis proposes a dual attention spatiotemporal fusion network composed of dual attention units and spatio-temporal attention fusion modules.The dual attention module(DAU)uses parallel spatial attention and channel attention to enhance features along the channel and spatial dimensions to distinguish different frame features,thereby suppressing the more useless and conveying useful features.The temporal-spatial fusion(TSA)module uses temporal attention and spatial attention to dynamically aggregate adjacent frames at the pixel level,so as to achieve a more effective fusion of inter-frame timing information and intra-frame spatial information.Compared with the other 7 advanced video super-resolution algorithms on the two test sets of Vid4 and Udm10,the algorithm in this thesis is more superior in both quantitative numerical results and qualitatively enlarged display images.