Yamada Yu, Okamoto Atsushi, Tomita Akihisa
Appl Opt. 2025 Mar 1;64(7):B6-B12. doi: 10.1364/AO.540689.
In the field of optics, a random phase mask (RPM) is used to efficiently encode and decode spatial complex amplitude distribution information of measurement targets into phase information. By encoding spatial complex amplitude distribution into a phase using an RPM, this distribution can be processed by modulating only the phase, which is efficient in computational terms. However, when encoding and decoding spatial complex amplitude distribution using an RPM, the resolution of optical devices such as a spatial light modulator (SLM) and charge-coupled device (CCD) becomes a bottleneck, resulting in decreased encoding and decoding accuracy. To address this issue, we propose a super-resolution method for phase images encoded with spatial complex amplitude distribution. This method uses a convolutional neural network (CNN) and a vision transformer (ViT), which are machine learning techniques widely used in computer vision. Through this super-resolution processing, we demonstrated that complex amplitude information can be encoded and decoded into phase images beyond the resolution of optical devices such as an SLM and CCD. Evaluation of the test images using peak signal-to-noise ratio (PSNR) showed improvements of 2.37 dB with the CNN and 1.86 dB with the ViT. Furthermore, we applied the proposed method to virtual phase conjugation based optical tomography (VPC-OT). The simulation results of measuring a microscopic target with a four-layer structure showed noise reduction at all depth positions and an improvement in the measurement accuracy of approximately 6-13 dB. (Details are shown in Fig. 7 and Table 2.) By applying the proposed method, measurement accuracy is improved with minimal computational operations, and without requiring additional optical systems or increasing the number of measurements. In addition, we examined the appropriate size of the machine learning model by observing the input image size (number of parameters) and loss progression.
在光学领域，随机相位掩模（RPM）用于将测量目标的空间复振幅分布信息高效地编码和解码为相位信息。通过使用RPM将空间复振幅分布编码为相位，可以仅通过调制相位来处理该分布，这在计算方面是高效的。然而，当使用RPM对空间复振幅分布进行编码和解码时，诸如空间光调制器（SLM）和电荷耦合器件（CCD）等光学器件的分辨率成为瓶颈，导致编码和解码精度降低。为了解决这个问题，我们提出了一种用于对用空间复振幅分布编码的相位图像进行超分辨率的方法。该方法使用卷积神经网络（CNN）和视觉Transformer（ViT），它们是计算机视觉中广泛使用的机器学习技术。通过这种超分辨率处理，我们证明了复振幅信息可以被编码和解码为超出SLM和CCD等光学器件分辨率的相位图像。使用峰值信噪比（PSNR）对测试图像进行评估表明，使用CNN时提高了2.37 dB，使用ViT时提高了1.86 dB。此外，我们将所提出的方法应用于基于虚拟相位共轭的光学层析成像（VPC-OT）。对具有四层结构的微观目标进行测量的模拟结果表明，在所有深度位置都有降噪效果，并且测量精度提高了约6-13 dB。（详细信息见图7和表2。）通过应用所提出的方法，在最小的计算操作下提高了测量精度，并且不需要额外的光学系统或增加测量次数。此外，我们通过观察输入图像大小（参数数量）和损失进展情况来研究机器学习模型的合适大小。
