基于多尺度特征融合网络的傅里叶叠层成像  

Fourier ptychography based on multi-scale feature fusion network

作　　者：宋东翰 王斌[1] 朱友强 刘鑫 SONG Dong-han;WANG Bin;ZHU You-qiang;LIU Xin(Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China;University of Chinese Academy of Sciences,Beijing 100049,China;Computer Vision and Pattern Recognition Laboratory,School of Engineering Science,Lappeenranta-Lahti University of Technology,Lahti 15210,Finland)

机构地区：[1]中国科学院长春光学精密机械与物理研究所,吉林长春130033 [2]中国科学院大学,北京100049 [3]拉彭兰塔-拉赫蒂理工大学工程科学学院计算机视觉与模式识别实验室,芬兰拉赫蒂15210

出　　处：《液晶与显示》2022年第11期1476-1487,共12页Chinese Journal of Liquid Crystals and Displays

基　　金：国家重点研发计划(No.2020YFA0714102)。

摘　　要：傅里叶叠层成像是一种实现光学系统高分辨率、大视场成像的技术。传统FP方法的高分辨率重建过程需要较高的孔径重叠率,导致采集图像数量较多,采样效率低。此外,FP重建算法的复杂度高,重建时间长。针对以上问题,本文结合深度学习,提出一种基于多尺度特征融合网络的傅里叶叠层成像算法,通过改进的特征金字塔卷积神经网络,能够从稀疏采样的低分辨振幅图像中提取特征信息并进行融合,实现超分辨的复图像重建。实验结果表明,在相同采样条件下,与传统方法相比,本文提出的深度学习算法提高了图像重建的质量,减少了90%以上的重建时间,并且对高斯噪声的鲁棒性较高。所提出的方法能够将相邻频域子孔径间的重叠率从50%降低至25%,减少50%的采集图像数量,大幅提高采样效率。Fourier Ptychography(FP) is a technology of achieving high-resolution,large field-of-view imaging of optical system. However,the high-resolution reconstruction based on traditional FP methods requires a high aperture overlap ratio,resulting in a large number of captured images and low sampling efficiency. In addition,the FP reconstruction algorithm has high complexity and long reconstruction time.Aiming at solving these problems of the FP,this paper proposes a deep learning algorithm based on multi-scale feature fusion network. Through the improved feature pyramid module,the feature information can be extracted from multiple low-resolution images captured by the FP imaging system,and the information is fused to achieve super-resolution reconstruction. Experimental results show that compared with traditional methods,the deep learning algorithm proposed in this paper improves the quality of image reconstruction,reduces the reconstruction time by 90%,and is more robust to Gaussian noise. In addition,the proposed method can reduce the overlap ratio between sub-apertures from 50% to 25% in frequency domain,and reduce the number of captured images by 50%,greatly improving the sampling efficiency.

关 键 词：计算成像 傅里叶叠层成像 特征金字塔 稠密连接 通道注意力 

分 类 号：O438[机械工程—光学工程] TP391[理学—光学]