| 高银,陈晨昕,李红云,郭霏霏,李俊.双域特征融合的Mamba去模糊方法[J].电子测量与仪器学报,2025,39(12):197-205 |
| 双域特征融合的Mamba去模糊方法 |
| Mamba deblurring method via dual-domain feature fusion |
| |
| DOI： |
| 中文关键词: 图像处理 计算装置 Mamba去模糊 双域特征融合 状态空间 |
| 英文关键词:image processing computing device Mamba deblurring dual-domain feature fusion state space |
| 基金项目:泉州职业技术大学开放课题(LERIS24-03)、国家自然科学基金(62001452)、福建省科技计划项目(2024T3040)资助 |
|
|
| Author | Institution |
| Gao Yin | Fujian Institute of Material Structure, Chinese Academy of Sciences, Quanzhou 362200，China |
| Chen Chenxin | 1.Fujian Institute of Material Structure, Chinese Academy of Sciences, Quanzhou 362200，China;
2.Fujian Normal University, Fuzhou 350117，China |
| Li Hongyun | Quanzhou Vocational and Technical University, Quanzhou 362200，China |
| Guo Feifei | Quanzhou Vocational and Technical University, Quanzhou 362200，China |
| Li Jun | 1.Fujian Institute of Material Structure, Chinese Academy of Sciences, Quanzhou 362200，China;
3.Quanzhou Vocational and Technical University, Quanzhou 362200，China |
|
| 摘要点击次数: 571 |
| 全文下载次数: 434 |
| 中文摘要: |
| 针对图像去模糊过程中单一域分析的局限和扫描特征分布差异化的问题，提出一种双域特征融合的Mamba去模糊方法。通过引入状态空间模型，同步提取模糊图像的空间结构特征与小波变换生成的多尺度频域特征，突破单一域分析的局限，实现空间域上下文信息与小波域高频细节特征在状态空间模型引导下的深度聚焦与自适应融合。设计双分支状态空间模块，分别独立建模空域与频域信息，精准适配空域结构特征与频域高频细节的差异化分布特性，在显著提升特征表征能力的同时，彻底规避扫描特征分布差异化，实现高质量的恢复。实验结果表明，所提方法在GoPro数据集上峰值信噪比（PSNR）达到33.75 dB，结构相似性（SSIM）为0.968；在HIDE数据集上PSNR为31.81 dB，SSIM为0.949；在RealBlur J和RealBlur R数据集上分别取得PSNR 32.92/0.937和40.15/0.974，显著优于对比方法。提出的方法在模糊去除、结构恢复、边缘保留和视觉效果方面的性能均优于经典去模糊方法，通过该方法设计出的装置能够在实际工程领域实现高精度清晰化处理。 |
| 英文摘要: |
| In view of the limitations of single-domain analysis and the differentiated distribution of scanning features in image deblurring, a novel Mamba deblurring method based on dual-domain feature fusion is proposed. By introducing a state-space model, the proposed method simultaneously extracts spatial structural features from blurred images and multi-scale frequency-domain features generated by wavelet transformation. This approach overcomes the constraints of single-domain analysis and enables deep integration and adaptive fusion of spatial-domain contextual information with high-frequency details in the wavelet domain, all under the guidance of the state-space model. A dual-branch state-space module is designed to independently model spatial and frequency-domain information, accurately adapting to the differentiated distribution characteristics of spatial structures and high-frequency details in the frequency domain. While significantly enhancing feature representation capabilities, the method effectively addresses the challenges posed by the differentiated distribution of scanning features and achieves high-quality image restoration. Experimental results demonstrate that the proposed method achieves PSNR of 33.75 dB and SSIM of 0.968 on the GoPro dataset, PSNR of 31.81 dB and SSIM of 0.949 on the HIDE dataset, and PSNR/SSIM of 32.92/0.937 and 40.15/0.974 on RealBlur-J and RealBlur-R datasets, respectively, outperforming classical deblurring approaches in terms of blur removal, structural restoration, edge preservation, and overall visual quality. Devices developed based on this method are capable of high-precision image enhancement in practical engineering applications. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
