Research On Light-Weight Image Super-Resolution Reconstruction Algorithm Based On Deep Learning

Image super-resolution is a typical low-level computer vision task,which aims to reconstruct the super-resolution image which is closer to the high-resolution image according to the low-resolution image which has lost some high-frequency information.Because high-resolution images contain more details,high-resolution images are extremely needed in Internet media,remote sensing,medical diagnosis,military reconnaissance,space exploration,and many other fields.However,the improvement of the imaging system at the hardware level to improve the resolution is limited by many factors,the algorithm for super-resolution reconstruction of low-resolution images at the software level has become an important method to improve the resolution of images.It also gives the image super-resolution reconstruction technology a high research significance.In recent years,image super-resolution reconstruction based on deep learning has shown outstanding performance in the field of image super-resolution reconstruction,but most of these algorithms achieve higher performance through deeper depth and a higher number of convolution kernels.As a result,most of the current methods have a large number of parameters and high computation.The computing power and storage space that can be provided in practical application scenarios such as in-orbit computing are often limited,so large models based on deep learning models are often difficult to deploy.To solve the above problems,the purpose of this thesis is to design a lightweight image super-resolution reconstruction algorithm based on deep learning with embedded deployment potential.For this purpose,a lightweight image super-resolution convolution neural network LRN(Lightweight Laplacian Pyramid Recursive and Residual Network)based on the Laplacian image pyramid is proposed in this thesis.By introducing the idea of parameter sharing and recursion,the network compresses the number of parameters to 3.98%of the number of parameters of its prototype algorithm Lap SRN(Laplacian Pyramid Super-Resolution Network).The super-resolution reconstruction experiments of 2x,4x,and 8x of the reference image datasets of LRN are carried out.The experimental results show that LRN achieves the same super-resolution reconstruction performance as Lap SRN with a very low number of parameters,which proves that LRN is a lightweight and effective algorithm Furthermore,the super-resolution reconstruction of Mars remote sensing image dataset proposed by LRN is tested,LRN shows a slightly higher objective evaluation index value than Lap SRN,and the two visual effects are similar.The experiment results show that LRN is effective for features of different domains,and the algorithm has a certain generalization ability.Although LRN maintains the network performance while reducing the number of parameters,due to the introduction of recursive blocks,the deepening of network depth,the computational complexity of LRN still needs to be reduced,and the performance needs to be improved.To solve this problem,this thesis proposes a lightweight image super-resolution reconstruction network LSTN(Light-weight super-resolution network using Swin Transformer)based on Swin Transformer.The experiment of super-resolution reconstruction of reference image data set of LSTN is carried out and the visualization analysis of lightweight degree is carried out.LSTN greatly improves the performance of the algorithm and reduces the amount of computation at the cost of a very low number of 28K103 parameters.The test results of Urban100 data set in the 2x super-resolution task improve 1.08d B and 0.012respectively compared with Lap SRN under PSNR and SSIM index.In terms of computation,compared with LRN,the amount of Multi-Adds of 2x super-resolution reconstruction of 720p image by LSTN is reduced from 374G(10~9)to 82G,which also shows an advanced level in comparison with other lightweight algorithms.The experimental results show that LSTN is a image super-resolution reconstruction algorithm based on deep learning with high performance,low number of parameters,low computational complexity and embedded deployment potential.Because LSTN has the potential to improve its performance,it further complicates LSTN and proposes a STN model to test the benchmark image data set of STN.Compared with Lap SRN,the test results of Urban100 dataset in 2x super-resolution task are improved by 2.27d B and 0.0229 respectively under PSNR and SSIM indicators.The experimental results prove the performance superiority of STN and the expansibility of LSTN network.