Extended depth-of-field resolution enhancement microscopy imaging for neutralizing the impact of mineral inhomogeneous surface

One of the most fundamental experimental methods in geoscience is to observe minerals under high magnification objectives. However, uneven microsurfaces in thin sections occur due to the irregular constituent distribution and varying hardness of minerals in natural rocks. Consequently, the conflict between large depth-of-field (DOF) and high-resolution in microscopy imaging leads to random out-of-focus issues when observing thin sections with high resolution microscopy. Although existing superresolution algorithms promise to improve visual performance, reconstructing images with both large DOF and high-resolution simultaneously remains challenging. We address this problem by guiding the networks with optical information. Utilizing DOF information from low-resolution data, we propose an optically induced generative adversarial network (OIGAN) to neutralize the impact through computational imaging. In OIGAN, optical DOF information from low-resolution data facilitates to achieve spatial-adaptive extended-DOF resolution enhancement imaging, without incorporating extended DOF high-resolution data for supervision. The approach, trained and evaluated on the dataset with 233,156 images (115,346 pairs of low- and high-resolution data), outperforms four comparison methods on various minerals and optical conditions, leading to at least 1.54dB increase on peak signal-to-noise ratio (PSNR). Specifically, OIGAN significantly improves the accuracy of fluid inclusion ice-melting temperature measurement, reducing mean error by 65%, and enhances mineral classification accuracy with 1.5%~15% increase. OIGAN offers an insight of integrating physical knowledge into neural networks, facilitating self-identification of minerals, automatic microthermometry of fluid inclusions and other geoscience tasks via microscopy.