基于多尺度局部与全局特征融合的轻量级超分辨率网络

Lightweight Super-Resolution Network Based on Multi-Scale Local and Global Feature Fusion

摘要: 当前图像超分辨率(SR)的研究重点方向是开发一种平衡图像重建性能和低资源消耗量的轻量级模型。Transformer擅长全局特征建模，但资源消耗量和参数量极高，同时欠缺探索精细局部细节的能力。而卷积神经网络(CNN)资源消耗低，但缺乏捕获全局特征的能力。为了解决这些问题，本文提出了一种基于纯CNN的模型，称为多尺度局部与全局特征融合网络(MLGFFN)。MLGFFN采用多尺度大核卷积块和局部特征提取块双分支分别捕获全部和局部特征信息，并且改进了特征融合方式。此外，MLGFFN还引入了特征增强前馈网络进一步细化前一级的特征图，对其中的重要特征信息进行强调，同时对冗余特征进行抑制。大量实验表明，本文提出的MLGFFN方法优于现有的轻量级SR模型，在轻量化设计和图像重建性能之间取得了良好的平衡。特别是在×4放大因子的Set14数据集上，MLGFFN的PSNR指标比SMSR模型高出0.21 dB，而参数量和FLOPs仅为SMSR的37.38%和40%。

Abstract: Current research in Image Super-Resolution (SR) focuses on developing lightweight models that achieve a balance between reconstruction performance and low resource consumption. While Transformer excels at global feature modeling, it suffers from excessive computational overhead and parameter counts, while lacking the ability to explore fine-grained local details. Conversely, Convolutional Neural Networks (CNNs) offer lower resource consumption but are deficient in capturing global features. To address these issues, this paper proposes a pure CNN-based model termed the Multi-scale Local and Global Feature Fusion Network (MLGFFN). MLGFFN utilizes a dual-branch architecture, comprising Multi-scale Large Kernel Convolution blocks and Local Feature Extraction blocks, to capture global and local feature information, respectively, while incorporating an improved feature fusion mechanism. Furthermore, a Feature Enhancement Feed-forward Network (FEFN) is introduced to further refine the feature maps from previous stages, emphasizing critical information while suppressing noise. Extensive experiments demonstrate that the proposed MLGFFN outperforms existing lightweight SR models, achieving a superior balance between lightweight design and image reconstruction performance. Notably, on the Set14 dataset with a ×4 upscaling factor, MLGFFN achieves a PSNR improvement of 0.21 dB over SMSR model, while utilizing only 37.38% of the parameters and 40% of the FLOPs required by SMSR.