Image Super-Resolution Reconstruction Algorithm Based On Deep Learning

With the rapid development of artificial intelligence and the Internet of Things(Io T),the combination of Io T and images is widely applied in various fields,such as smart cities,smart healthcare and smart displays.The ensuing phenomenon is that people’s requirements for image information are getting higher and higher,and clear images with high resolution are favored.However,the images obtained from the actual scene are often low-resolution and blurred due to the interference of the acquisition equipment and the external environment.For this reason,image super-resolution reconstruction techniques were developed,which aim to reconstruct high-resolution images from low-resolution images.Although existing super-resolution reconstruction algorithms have achieved significant improvements in performance and efficiency,they are still challenging to apply in real-world scenarios.Firstly,this is because the degradation pattern of the actual scene image is unknown,and the existing methods are difficult to adapt to the changes in the image degradation space,resulting in artifacts and line distortion in the recovered image.Secondly,the computational complexity of the existing methods is too high to strike a balance between accuracy and speed,which leads to difficulties in terminal deployment.In response to the above problems,this paper mainly explores two aspects of the image feature domain and image degradation domain,and the main research contents are as follows:(1)Since the existing state-of-the-art networks have a poor ability to fuse features where scales are inconsistent,and they cannot flexibly balance the dependence between global-local information.Therefore,this paper proposes a super-resolution network based on global-local attention-guided to reconstruct high-quality images.The network consists of a multiscale generator and a Transformer discriminator.In particular,a weighted multi-scale feature aggregation module based on parameter self-learning is designed,which captures information at different scales by using multiscale convolution and Swin Transformer in parallel,and guides the effective fusion of global-local information through the attention mechanisms and weighted learning factors.In addition,this paper design a new hybrid loss,which utilizes weight learning and regularization penalty,to enhance the global-local information recovery ability of the image.This avoids the problems of creating local artifacts and blurring the boundary of the restored image.(2)Since current lightweight models either lead to loss of accuracy or incur additional overhead,it is difficult to apply them in real scenarios.This paper proposes a lightweight superresolution reconstruction framework based on the soft weighting of pixels,in order to reduce the number of model parameters and inference time while obtaining high performance.In particular,a new feature-guided block(FGB)is designed for receptive field enhancement to enhance high-frequency details by focusing on the pixel position information.In addition,this paper customizes an enhanced self-residual block(ESRB)to achieve efficient inference while deepening feature learning.The new framework proposed in this paper can greatly reduce memory consumption and speed up inference while satisfying high accuracy.(3)Since the premise of the high performance of the super-resolution reconstruction algorithm is to estimate an accurate degradation model,and the existing degradation models are still insufficient in dealing with the complexity of real image degradation,they are difficult for them to mimic real degradation modes with multiple directions and types.Therefore,this paper proposes a new image degradation model,namely the domain randomization degradation model,which uses a randomized degradation model for different regions of the image to accommodate the degraded domains of real-world images,which can enrich the sample diversity and simulate the real-world image degradation as much as possible.In addition,this paper is equipped with an efficient network for fine-grained feature fusion.This network can enhance the local information of the reconstructed image while suppressing the global noise,and finally reconstructing a high-quality image.In this paper,experimental simulations are performed on several publicly available datasets,including classical image data,compressed image data,and real-world image data.Experimental results show that the proposed method in this paper performs well in subjective visualization,recovers realistic image textures and reduces boundary artifacts compared with existing methods.In addition,the method in this paper reduces the complexity of the model,achieves a certain balance between performance and parameters,and also achieves good results in the evaluation of objective metrics,which proves that the proposed method has good generalization.