This study introduces a self-supervised learning algorithm, named SSL-Depth, which achieves a 16-fold speed enhancement in 3D surface imaging based on confocal microscopy.
The field of 3D surface imaging holds significant importance in scientific and industrial research, offering detailed insights into surface properties. Nevertheless, traditional 3D imaging techniques necessitate the acquisition of substantial data and exhibit slower measurement speeds, thereby limiting their applicability in dynamic measurement applications.
In order to address the aforementioned issues, Professor Sun Fangwen and his team from the University of Science and Technology of China designed the SSL-Depth algorithm. The algorithm markedly improves the efficiency and precision of confocal microscopy measurements through a self-supervised learning approach. The research results are published in Chinese Optics Letters, Vol. 22, Issue 6, 2024.
.png)
Fig. 1. Network architecture. The encoder consists of a Vision Transformer (ViT) and an adapter for accelerated training. Intensity, depth, and width are predicted by three decoders to finally reconstruct the raw data.
The SSL-Depth algorithm exploits the benefits of self-supervised learning, which does not rely on labelled data or other pre-established rules. In contrast, it learns effective representations of the data by making more comprehensive use of the measurement data, thus avoiding the dependency on labeled datasets that is typical of traditional supervised learning. This method employs a physics-informed Transformer architecture that is capable of efficiently predicting intensity, depth, and PSF peak width directly from raw images acquired by confocal microscopy. This results in a reduction of the Z-axis sampling frequency, which in turn yields a notable increase in speed. The research team conducted practical experiments to assess the performance of SSL-Depth. The results demonstrated that SSL-Depth not only maintained high imaging quality but also increased measurement speed by 16 times. This highlights the potential of SSL-Depth in 3D surface imaging with confocal microscopy.
The research team conducted a series of experiments using a commercial confocal microscope on surface roughness comparators to validate the effectiveness of the SSL-Depth algorithm. The experimental results demonstrated that SSL-Depth performed exceedingly well in 1x, 4x, and 16x modes in comparison to traditional commercial algorithms. In particular, SSL-Depth achieved speeds and accuracy in 16x mode that traditional algorithms were unable to match.
The team has applied the SSL-Depth method to bullet 3D topography, successfully reducing the measurement time of bullets by six times, significantly enhancing the efficiency of firearm and bullet archiving for law enforcement. In the future, the research team plans to extend this method to the field of quantum sensing based on confocal microscopy and NV centers, aiming to improve the precision, sensitivity, and speed of micro-nano electromagnetic measurements.
