| 引用本文: | 傅慧滢,王 瑜,杨高明.基于密集高效 Transformer 的图像超分辨率重建(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2026，43（1）：72-79 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| :目的 针对超分辨率网络存在参数量大和图像特征提取能力不足导致重建图片质量不佳的问题,提出了一
种新的图像超分辨率模型 LHNet。 方法 在深层特征提取模块部分,设计对称密集 CNN,以对称结构连接卷积模
块,加强模型对图像局部信息的提取能力;引入双注意力模块,使网络关注像素和空间信息;此外,使用密集连接方
式连接高效 Transformer 模块,增强模型全局特征的提取能力,从而更好地恢复图像纹理细节。 结果 实验结果在
Set5、Set14、BSD100、Urban100 和 Manga109 等五个方法数据集进行测试。 结果表明:在 Set5 数据集上,放大倍数为
3 时,LHNet 方法相对于 IMDN 方法减少了 99 K 的参数量,同时 PSNR 值提高了 0. 13 dB,SSIM 提高了 0. 000 9。 与
当前其他方法,如 LAPAR-A、ShuffleMixer 等方法相比,LHNet 方法也表现出优越的性能。 结论 LHNet 方法可以在
使用相对较少参数量的同时提高重建图片的质量,从而实现参数量和性能之间的平衡。 |
| 关键词: 超分辨率重建 对称结构 双注意力模块 密集连接 高效 Transformer |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Image Super-Resolution Reconstruction Based on Dense and Efficient Transformer |
|
FU Huiying WANG Yu YANG Gaoming
|
|
School of Computer Science and Engineering Anhui University of Science and Technology Huainan 232001 Anhui
China
|
| Abstract: |
| Objective To address the problems of a large number of parameters and insufficient image feature extraction
ability in super-resolution networks which lead to poor quality of reconstructed images a new image super-resolution
model LHNet is proposed. Methods In the deep feature extraction module a symmetric dense CNN was designed. The
convolution module was connected in a symmetric structure to enhance the model?? s ability to extract local image
information. A dual attention module was introduced to make the network focus on pixel and spatial information. In
addition the efficient Transformer modules were connected using a dense connection method to strengthen the model?? s
global feature extraction ability thus better restoring the image texture details. Results The experimental results were
tested on five datasets namely Set5 Set14 BSD100 Urban100 and Manga109. The results showed that on the Set5
dataset when the magnification factor was 3 the LHNet method reduced the number of parameters by 99 K compared with
the IMDN method. Meanwhile the PSNR value increased by 0. 13 dB and the SSIM increased by 0. 000 9. Compared
with other current methods such as LAPAR-A and ShuffleMixer the LHNet method also showed superior performance.
Conclusion The LHNet method can improve the quality of reconstructed images while using a relatively small number of
parameters thereby achieving a balance between the number of parameters and performance. |
| Key words: super-resolution reconstruction symmetric structure dual attention module dense connection efficient
Transformer |
