Research On Single Image Super Resolution Based On Deep Learning

High-resolution images are a valuable source of detailed texture information essential for numerous computer vision tasks.However,obtaining high-resolution images pases challenges due to the high cost of image acquisition,limited video transmission bandwidth,and digital processing limitations,While hardware-based solutions for increasing image resolution can be costly and restrictive,software-based super-resolution reconstruction methods using deep learning offer flexibility,scalability,and adaptability to various scenarios and evolving requirement.Among these methods,deep learning-based super-resolution reconstruction using convolutional neural networks has garnered considerable attention in recent years.This is primarily due to its capability to learn nonlinear mappings between low-resolution and highresolution images,enabling the generation of images with enhanced details and clear textures,less constrained by application scenarios.Thus,this paper aims to construct an efficient superrcsolution reconstruction model based on deep learning.The specifie objectives are as follows:1)Addressing the chailenge of utilizing the scale and channel features of low-rcsolution images effectively,which is crucial for designing a high-quality super-resolution reconstrucetion model.To tackle this issue,we propose a multi-scale fractal residual attention network.Firstly,a multi-sale extension rule is devised to form a fraclal residiual block with a multi-path structure through successive extensions.This block is employed to detect the multi-scale features of lowresolution images,Furthermore,the base block within the fractal residual block incorporates the proposed combined dilated conolution and dual enhanced channel attention.This combination facilitates information interaction within and across channcls trough channel attention and onedimensional convolution while enhancing grouping through channel shuffling.Lastly.low-level spatial feature and high-level features are fused using global residuals and feature fusion.The proposed module’s effectiveness is evaluated through quantitative assessments,ablation experiments,and comparison experiments conducted on five benchmark datasets.2)The Transformer model can better capture the global dependencies in the sequence through the self-attention mechanism,but it faces the problems of insufficient local feature extraction and high computational resource consumption when processing images.The CNN model can extract local features more effectively with lower computational complexity and memory consumption through the sliding window mechanism and parameter sharing feature,but it is weak in sensing long-range location information of images.To solve these problems,this paper proposes a superresolution reconstruction method based on a large kernel separation convolutional feature distillation network,which uses a multi-path structure to learn different horizontal feature representations.In addition,this paper designs a multi-scale large kernel separable convolutional block,which considers the powerful global information capturing ability of self-attention mechanism and the powerful local receptive ability of convolution and can better extract global features and local features.Meanwhile,this paper also uses lightweight normalized channel attention,which further enhances the model performance while realizing the lightweight design of the network model.Finally,this paper uses four evaluation metrics for quantitative evaluation in five benchmark datasets,the experimental results show that the proposed methods achieve competitive results.