In computer vision and robotics, accurate 3D positioning and trajectory determination are crucial for a variety of applications, including industrial and clinical [1]. Neural networks, including convolutional neural networks (CNNs) or Vision Transformers (ViT) play a significant role in visual data processing [2]. Digital holography (DH) in microscopy enhances the analysis of object amplitude and phase in a single image with off-axis configuration, improving the accuracy of in-focus position detection without mechanical adjustments. Combining Deep Neural Networks (DNN), mixing version of the GedankenNet model [3] and a UNet-like model [4] with DH provides a promising solution for accurately controlling complex trajectories of micro-objects in automated microscopy in real-time constrains [5].

DNNs inspired by biological neural networks, process, classify, and predict complex data through multi-layer structures. These networks employ non-linear transformations from input to output layers, enabling tasks like linearization in higher-dimensional spaces [4]. Optimization of DNN results involves a learning step, training the network with input-output data pairs. Adequate training data volume is crucial for optimal performance. DNNs, notably convolutional CNNs and ViT models, have demonstrated high effectiveness in tasks like image classification, computer vision, and solving complex problems such as autofocusing in DH [2, 3].

DH is an advanced imaging technique capturing both amplitude and phase of an object’s entire wavefield using a CMOS imaging sensor. In Figure 1, we show typical experimental digital hologram a 2D pseudo-periodic pattern as phase object to perform 3D pose control in 3D through a microscope [2]. This study explores DH coupled with a computer micro-vision approach, employing phase correlation image processing techniques for sub-voxel sample pose measurements in micro-robotics [6, 7].

Digital hologram reconstruction relies on the Angular Spectrum Method [8], and a Lyncee‐Tec Digital Holographic Microscope (DHM) equipped with 10× MO lens, adapts these principles to micro-objects, see reference [2] for experimental details. DHM works with digital autofocusing, enables automated microscopy and 3D pose control of micro-objects. Recent research highlights the use of DNN for faster auto-autofocusing in DHM through statistical image reconstruction, treating autofocusing as a classification or regression task [5]. The challenges include improving multiscale sensitivity for automated microscopy in 6 degrees of freedom (DoF) pose estimation while maintaining a broad field of view and depth of field [1]. A 2D pseudo-periodic pattern serves as a referencing sample (Fig. 1(c) and (d)). High-tech micro-assembly platforms in robotics demand translation and rotation stages (Fig. 1(a)), addressing increasingly complex tasks with nanoscale positioning resolution and large-scale movements beyond the centimetre range. This work addresses the challenge to target 3D inference and video-rate control of samples for complex micro-nano manipulation such as 3D MEMS micro-nano-assembly and alignment, 3D nanoprinting, visual servoing for 3D nanopositioning [1].

In this work, we combine previous autofocusing with DHM accelerated with DNN [2] giving Z position and a new approach to determine in the same time X and Y coordinates. In Figure 2(a–c), the structure of the XY Model (consisting of a series of 2D Convolution Layers and Max Pooling Layers) based on the UNet architecture [4] is presented, specifically designed for 3D pose estimation. The model takes a Region of Interest (ROI) extracted from the input hologram, initially sized at 768 × 768 pixels within a hologram of 1024 × 1024 pixels. The resulting output from the model is a reconstructed thumbnail of 64 × 64-pixels, encapsulating the X and Y positional information [6]. Subsequently, Figure 2(d–f) outlines the arrangement of the Z Model, which is based on an adapted version of a GedankenNet model proposed in [3]. The primary distinctions from the original version are that it accepts a single image as input and the input size has been minimized to 128 × 128 pixels for faster computation of the Spectral Conv2D Layers (Fig. 2(f)). The XY Model’s uniqueness lies in not reconstructing an image of the same size as the input (Fig. 2(b) depicts the initial Conv2D layers downsizing the input to 64 × 64).

We address this issue by applying DNNs to micro-vision measurement of 3D trajectories with DH. Recently, we demonstrated the ability of new generation of deep neural networks such as ViT to predict the in-focus distance with a high accuracy [2]. In a previous work, we also showed the ability of 2D pseudo-periodic pattern combined to conventional imaging system, used as in-plane position encoder, has allowed a 108 range-to resolution ratio through robust phase-based decoding [7]. Here, we present DNNs dedicated to hybrid approach combining computer micro-vision and DHM, able to perform simultaneously in-plane and out-plane measurements, at video-rate and without in focus full image reconstruction. The experimental setup is presented in Figure 1. It consists in a DHM, a hexapod capable of precise motions along the 6DoF and a micro encoded pattern. We also show a typical hologram obtained and its reconstruction (Fig. 1(b)). The interferometric character of DH converts out-of-plane position of the sample in phase data that, combined with in-plane information retrieved from the micro-structured pattern, allows accurate measurement of 3D trajectories. DNNs speed up data processing and infer video-rate position detection.

DNNs require training to realize expected tasks and to reach the best performances. In our work, the training step is conducted from a dataset constituted by simulated holograms. Various experimental parameters have been considered in simulations such as spherical aberration introduced by objective microscope lens, and has been implemented in simulated hologram datasets, with the aim of being able to mimic real experimental conditions. To rigorously evaluate the effectiveness of the proposed methodology, which integrates DH with DNNs and video-rate micro-vision, we conducted a comprehensive validation through simulation. Our primary objective was to assess the DNNs capability to predict a simulated 3D trajectory under precisely controlled conditions. For this purpose, we selected a Lissajous’ figure (result of superposing two harmonic motions on the X-Y plane). This complex trajectory served as a challenging yet well-defined path for rigorously testing the capabilities of the DH-DNN system. We simulated a complete 3D trajectory of 2D pseudo-periodic pattern with period of 9 μm, displaced by the hexapod stage (Fig. 1(a)), along the two-dimensional Lissajous trajectory in the X-Y plane and generated corresponding sequence of digital holograms. This trajectory was then extended into the third dimension by introducing incremental steps along the Z-axis, simulating motion in depth. Each step in the Z-direction corresponds to a subsequent holographic reconstruction distance for the simulated hologram. Subsequently, the generated holographic datasets were used in DNNs for training step and infer the trajectory. The networks were tasked with accurately predicting the Lissajous’ trajectory based on the holographic dataset inputs, essentially capturing and replicating the complex curve in their predictions. To analyse each hologram (inference mode), both models are used (Fig. 2), XY Model and Z Model to get the associated thumbnail and Z distance. A post-processing algorithm is applied on the reconstructed thumbnail to extract the binary vectors representing the positions (X and Y) (Fig. 2c). To convert the binary vectors into meaningful micron-scale coordinates, each vector within the complete sequence of bits is identified. Those indexes are used to compute the final X and Y coordinates as described in [6].

We present the results obtained from the DH-DNN system methodology for predicting 3D trajectories. The models (XY Model and Z Model) have been trained using a total of 65000 simulated holograms. The XY Model is using binary cross entropy loss. The Z Model has been trained using a cross-validation method using the TanhExp loss function [9]. Both models are trained using the Adam optimizer. The models have been tested on a simulated trajectory of 1121 holograms. In Figure 3(a), the list of outliers (red points), the simulated (dashed blue line) and estimated (green line) trajectories are shown in 3D space. The accuracy exceeds 98% which demonstrates the system’s ability to correctly estimate the 3D poses. Figure 3(b) provides a visual representation of the error along the Z axis and the deviation on the X-Y plane (L2-norm). This graphical depicts the precision of DNN predictions, revealing a max error of 25 μm on X-Y and less than 1 μm on Z. This X, Y level of performance must be compared with a maximum encoded area of 11 × 11 cm². This allows video-rate monitoring of large displacements with a coarse but sufficient accuracy whereas eventual fine 3D pose is controlled by high accurate but much slower conventional processing.

Figure 4 shows the matching rate associated to each estimated 3D pose. This underscores that a rate level between 90 and 100 is adequate for accurately decoding the correct position. The precision along the Z axis is of the same magnitude as in [2]. These results emphasize the DH-DNN methodology’s capability to provide highly accurate and detailed predictions of three-dimensional trajectories. This highlights its practical utility in real-time micro-robotics and micro-vision applications. Moreover, the average inference speed is below 20 ms on a NVidia RTX 3090 32 GB mainly consumed by the data transfer of the images to the GPU (XY Model: 7.5 ms inference; Z Model: 2.5 ms inference; 10 ms for the data transfer).

We propose a method that enables the direct determination of 3D positions from hologram space with a mean error of 1 μm on Z and 12 μm on X-Y, effectively bypassing the need for full holographic image reconstruction. These errors must be compared to the complete encoded area of 11 × 11 cm². Moreover, our study offers a thorough analysis of the matching rate levels attributed to each 3D pose. We believe it is the first time a GedankenNet model is used as a regression tool. The modified GedankenNet (Z Model) achieved an inference speed of 2.5 ms, contrasting with the over 20 ms required by a TViT [2].