Deep-learning-supported depth-from-focus microscope

Kurzfassung

Compact sensor systems, such as those used in rover missions, can benefit from 3D measurement techniques if they are accurate and resource-efficient. The depth from focus (DFF) method derives depth information solely from variations in focus and requires no additional hardware beyond a camera system. Compared to classic approaches, deep depth from focus (DDFF) methods using convolutional neural networks (CNNs) can reduce the number of images needed for depth estimation. This lowers both energy consumption and memory requirements. In this work, a microscope setup was upgraded with a DDFF capability, demonstrating how DDFF can be applied to the measurement of geological samples. Experimental results indicate that DDFF outperforms conventional DFF methods in depth accuracy under limited image acquisition conditions. This study highlights the potential of DDFF as a powerful tool for in-situ scientific analysis in space exploration and other compact sensing applications.