卷积和自注意力融合的单图像超分辨率网络
Convolution and Self-Attention Fusion for Single-Image Super-Resolution Network
马勇1
作者信息
- 1. 福州大学先进制造学院,福建 泉州 362200
- 折叠
摘要
近年来,超分辨率重建已经成为图像处理领域的一个研究热点.然而,超分辨率重建面临着诸多挑战,当模型参数过大时,虽然能取得良好的性能,但需要巨大的内存成本.针对目前大多数图像超分辨率网络无法做到既能实现良好的性能,又能保持网络模型轻量级的问题,提出了一种用于单图像超分辨率的新型轻量级双阶段网络.具体来说,设计了一种轻量级卷积模块用于局部特征提取,同时引入了一种轻量级Transformer模块学习图像的长期依赖关系,用于建模全局信息.实验结果表明,所提模型在客观评价指标和视觉效果上均表现良好.
Abstract
In recent years,super-resolution reconstruction has become a research hotspot in the field of image processing.However,super resolution reconstruction faces many challenges.When the model parameters are too large,although it can achieve good performance,it needs huge memory cost.Aiming at the problem that most image super-resolution networks can not achieve good performance and keep the network model lightweight,this paper proposes a new lightweight two-stage network for single image super-resolution.Specifically,a lightweight convolution module is designed for local feature extraction,while a lightweight Transformer module is introduced to learn long-term image dependencies for modeling global information.The experimental results show that the proposed model performs well in both objective evaluation index and visual effect.
关键词
图像超分辨率/轻量级双阶段网络(LDSNet)/Transformer模块/卷积神经网络(CNN)/自注意力Key words
image super-resolution/Lightweight Dual-Stage Network(LDSNet)/Transformer module/Convolutional Neural Network(CNN)/self-attention引用本文复制引用
出版年
2024
NETL
NSTL
万方数据