J Opt Soc Am A Opt Image Sci Vis. 2021 Oct 1;38(10):1570-1580. doi: 10.1364/JOSAA.424271.
Digital holography is a useful tool to image microscopic particles. Reconstructed holograms give high-resolution shape information that can be used to identify the types of particles. However, the process of reconstructing holograms is computationally intensive and cannot easily keep up with the rate of data acquisition on low-power sensor platforms. In this work, we explore the possibility of performing object clustering on holograms that have not been reconstructed, i.e., images of raw interference patterns, using the latent representations of a deep-learning autoencoder and a self-organizing mapping network in a fully unsupervised manner. We demonstrate this concept on synthetically generated holograms of different shapes, where clustering of raw holograms achieves an accuracy of 94.4%. This is comparable to the 97.4% accuracy achieved using the reconstructed holograms of the same targets. Directly clustering raw holograms takes less than 0.1 s per image using a low-power CPU board. This represents a three-order of magnitude reduction in processing time compared to clustering of reconstructed holograms and makes it possible to interpret targets in real time on low-power sensor platforms. Experiments on real holograms demonstrate significant gains in clustering accuracy through the use of synthetic holograms to train models. Clustering accuracy increased from 47.1% when the models were trained only on the real raw holograms, to 64.1% when the models were entirely trained on the synthetic raw holograms, and further increased to 75.9% when models were trained on the both synthetic and real datasets using transfer learning. These results are broadly comparable to those achieved when reconstructed holograms are used, where the highest accuracy of 70% achieved when clustering raw holograms outperforms the highest accuracy achieved when clustering reconstructed holograms by a significant margin for our datasets.
数字全息术是一种用于对微观粒子成像的有用工具。重建的全息图提供了高分辨率的形状信息，可用于识别粒子的类型。然而，重建全息图的过程计算量很大，并且不容易跟上低功率传感器平台的数据采集速度。在这项工作中，我们探索了使用深度学习自动编码器和自组织映射网络的潜在表示以完全无监督的方式对未重建的全息图（即原始干涉图案的图像）进行物体聚类的可能性。我们在不同形状的合成全息图上证明了这一概念，其中原始全息图的聚类达到了 94.4%的准确性。这与使用相同目标的重建全息图达到的 97.4%的准确性相当。使用低功耗 CPU 板直接对原始全息图进行聚类，每张图像的处理时间不到 0.1 秒。与重建全息图的聚类相比，这代表处理时间减少了三个数量级，使得在低功率传感器平台上实时解释目标成为可能。在真实全息图上的实验表明，通过使用合成全息图来训练模型，可以显著提高聚类精度。当仅在真实原始全息图上训练模型时，聚类精度从 47.1%提高到 64.1%，当完全在合成原始全息图上训练模型时，聚类精度进一步提高到 75.9%，当使用迁移学习在合成和真实数据集上训练模型时，聚类精度提高到 75.9%。这些结果与使用重建全息图时的结果大致相当，在我们的数据集上，当聚类原始全息图时达到的最高精度 70%明显优于聚类重建全息图时达到的最高精度。
