【作者】 陈旭；

【导师】 汪荣贵； 范联伟；

【作者基本信息】 合肥工业大学 ， 计算机技术（专业学位）， 2025， 硕士

【摘要】 图像超分辨率技术旨在从低分辨率图像中重建高分辨率细节,在移动端视觉、医学成像和视频增强等领域具有重要应用价值。然而,传统深度超分辨率模型存在计算复杂度高、参数量大等问题,难以部署于资源受限的边缘设备。本文聚焦轻量级图像超分辨率算法,围绕模型效率与重建质量的平衡展开研究。目前的主流图像超分辨率模型采用深度学习技术,具体包括卷积神经网络、生成对抗网络、注意力机制等,本文主要采用深度学习技术进行网络构建,然后对模型进行轻量化设计,最后在主流的数据集上测试性能相关指标,验证所出方法的有效性。主要研究内容如下:(1)针对传统超分模型局部边缘特征学习不充分问题,同时升模型的长距离依赖建模能力,本文出一种基于信息增强的轻量级Transformer超分辨率网络,特殊设计的高频取模块可以很好地取增强后的图像中的高频信息,另外在Transformer分支中使用多方向窗口计算注意力可以充分学习图像中的潜在特征和细节信息,高图像的重建质量。在注意力计算部分,本文出的动态参数共享机制能够根据局部图像特征进行自适应参数共享,从而大大降低模型的参数量。经过实验验证,本文出的方法能够以较低的参数量和内存占用取得优异的超分辨率性能。(2)为进一步降低复杂的自注意力计算带来的性能消耗,本文出一种创新的轻量级跨域混合注意力网络用于图像超分辨率重建。网络通过频域空间域决策网络动态分析输入图像在频域与空间域的特征重要性,生成混合权重矩阵指导后续特征融合。设计双分支混合注意力模块,结合频带注意力、空间注意力与通道注意力机制,实现频域全局结构恢复与空间域局部细节增强的协同优化。该网络采用轻量化结构设计,通过残差连接升训练稳定性,在降低参数量的同时保持优异的细节恢复能力。更多还原

【Abstract】 Image super-resolution technology aims to reconstruct high-resolution details from low-resolution images,holding significant application value in mobile vision,medi-cal imaging,and video enhancement.However,traditional deep super-resolution mod-els suffer from high computational complexity and excessive parameter counts,making them difficult to deploy on resource-constrained edge devices.This paper focuses on lightweight image super-resolution algorithms,exploring the balance between model ef-ficiency and reconstruction quality.Current mainstream image super-resolution models employ deep learning techniques, including CNNs,GANs and attention mechanisms.This work primarily adopts deep learning for network construction,followed by lightweight model design,and finally evaluates performance metrics on mainstream datasets to validate the effectiveness of the proposed methods.The main contributions are as follows:(1)To address insufficient local edge feature learning in traditional super-resolution models and enhance long-range dependency modeling,we propose an information-enhanced lightweight Transformer-based super-resolution network.A specially designed high-frequency extraction module effectively captures enhanced high-frequency infor-mation from images.Additionally,a multi-directional window-based attention mecha-nism in the Transformer branch comprehensively learns latent features and fine details,improving reconstruction quality.The proposed dynamic parameter sharing mechanism in attention computation adaptively shares parameters based on local image characteris-tics,significantly reducing model complexity.Experimental results demonstrate that our method achieves superior super-resolution performance with fewer parameters and lower memory consumption.(2)To further mitigate the computational overhead of complex self-attention mecha-nisms,we propose an innovative lightweight cross-domain hybrid attention network for image super-resolution.The network dynamically analyzes feature importance in the fre-quency and spatial domains through a frequency-spatial domain decision network,gen-erating hybrid weight matrices to guide feature fusion.A dual-branch hybrid attention module combines frequency-band attention,spatial attention,and channel attention mech-anisms,enabling synergistic optimization of global structural recovery in the frequency domain and local detail enhancement in the spatial domain.The lightweight architecture employs residual connections to enhance training stability while maintaining outstanding detail restoration capabilities with reduced parameters.更多还原