| 摘要: |
| 合成孔径光学成像系统中由于通光面积不足和平移误差等因素，最终成像易出现退化模糊的现象，基于数学模型的传统复原方法难适用于不同系统间。提出一种基于局部-全局特征增强Transformer的网络，解决了合成孔径光学系统中高分辨率退化图像难以复原的问题。提出基于门控机制的残差卷积层，利用可变形卷积和简单门控机制关注图像的局部特征；构造基于线性注意力与门控机制的Transformer层，在减少计算复杂度的同时建立图像信息之间的长距离依赖关系；针对合成孔径光学成像系统中产生的振铃现象，提出自适应尺度特征增强模块，对不同尺度下的特征利用特征权重进行二次学习，增强特征中结构信息的锐化表达能力，避免在复原过程中被振铃现象产生的伪影干扰。实验结果表明，该网络在计算复杂度减少了37.85%的同时，峰值信噪比和结构相似度相比于其余方法平均提高了8.07%和3.17%，能有效复原合成孔径光学系统中高分辨率退化图像。 |
| 关键词: 合成孔径光学图像 图像复原 局部-全局特征增强 |
| DOI：10.20079/j.issn.1001-893x.240903005 |
|
| 基金项目:国家自然科学基金资助项目(62071240，62106111)；2024年江苏省研究生创新项目(SJCX24_0446)；无锡市科技发展资金“太湖之光”科技攻关项目(K20231004) |
|
| Image Restoration Method for Transformer Synthetic Aperture Optical System with Local-Global Feature Enhancement |
| TONG Junyi,ZHANG Yinsheng |
| (1.School of Electronic and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China;2.Jiangsu Engineering Research Center for Sensor Fusion Photonic Devices and System Integration,Wuxi University,Wuxi 214105,China) |
| Abstract: |
| In the synthetic aperture optical imaging system, due to insufficient light transmission area and translation errors, the final imaging is prone to exhibit degraded blurring. Traditional restoration methods based on mathematical models are difficult to be applied across different systems. A network based on local-global feature-enhancing Transformer is proposed to solve the problem of difficult restoration of high-resolution degraded images in synthetic aperture optical systems. A residual convolution layer based on gated mechanism is proposed, which utilizes deformable convolution and simple gated mechanism to focus on the local features of the image. A Transformer layer based on linear attention and gated mechanism is constructed to reduce the computational complexity while establishing long-distance dependency relationships between image information. For the ringing phenomenon generated in synthetic aperture optical imaging systems, an adaptive scale feature enhancement module is proposed. For features at different scales, secondary learning is conducted using feature weights, enhancing the sharpness expression ability of structural information in the features and avoiding interference from artifacts generated by the ringing phenomenon during the restoration process. Experimental results show that this network reduces the computational complexity by 37.85% while the peak signal-to-noise ratio and structural similarity are on average improved by 8.07% and 3.17% compared with those of other methods, and can effectively restore high-resolution degraded images in synthetic aperture optical systems. |
| Key words: synthetic aperture optical image image restoration local-global feature enhancement |
《电讯技术》编辑部 