Research On Lightweight Super-Resolution Algorithm Based On Grouping Fusion Of Feature Frequencies

Single image super-resolution algorithm is one of the important tasks in the field of computer vision.Its goal is to restore high-resolution images from images which are low-resolution.At present,deep convolutional neural networks have significantly improved the performance of single image super-resolution,and have become the mainstream algorithms.Ordinarily,the larger the network size,the better the performance,but the higher the computing cost and storage resources are required,which limits their applications on resource-constrained devices.Therefore,it is very necessary to study the design of lightweight super-resolution network models.In this paper,some lightweight super-resolution networks that balance model complexity and performance are proposed by improving the network structure.The main research work and results of this thesis are summarized below:(1)We proposed an image super-resolution algorithm via grouping fusion of feature frequencies.Firstly,the residual concatenation block is constructed,which uses both residual learning and channel concatenation to promote the transfer and fusion of local feature information;Then,the proposed multi-way hybrid attention block is used to collect different feature information clues from different scales to improve the detail fidelity of the network,providing rich image information and more high-frequency details for network reconstruction;Finally,the feature frequency grouping fusion block is used to fuse global feature information.The feature frequency grouping fusion block utilizes the differences between the high and low frequency information of features,firstly groups the features,and then gradually fuses the features of each group starting from the group with the largest feature difference.Experimental results demonstrate that the method in this paper can make better use of the hierarchical information in the network make better fusion of high-frequency and low-frequency feature information,and improve the reconstruction quality of super-resolution images.(2)We proposed a super-resolution algorithm based on lightweight convolution.First,a novel group convolution is designed,called progressive interactive group convolution,which is more efficient than traditional group convolution.It first separates the input features,and progressively processes the separated features,and then progressively interacts the information between the processed group and the unprocessed group,thereby enhancing the interaction between features on different channels;Then,a lightweight convolutional block dedicated to image super-resolution is constructed,which consists of three parts:progressive interactive group convolution,channel shuffle,and depthwise separable convolution,which effectively reduces the network parameters;Finally,a lightweight residual concatenation block is designed as the basic building block of the network,which makes the network more lightweight and competitive.In order to better fuse multi-level feature information and extract richer details,this method introduces the feature frequency grouping fusion block and multi-way hybrid attention block proposed above.Experimental results demonstrate that the algorithm we proposed balances the complexity and performance well,and is a very practical method,which can improve the quality of the reconstructed image.In summary,the algorithms proposed in this thesis aim to balance the complexity and performance of the model,build lightweight super-resolution networks,and improve the quality of reconstructed images.Under the same dataset conditions,this paper conducted comparative experiments with the reconstruction results of lightweight algorithms with similar parameter quantities.The results showed that the proposed algorithms achieved good performance in both subjective visual quality and objective quantitative metrics for super-resolution reconstruction.The algorithms in this thesis provide effective methods to design lightweight super-resolution algorithms,providing a new option for super-resolution image processing on resource constrained devices.