Instance segmentation from particle holograms

Abstract

This study introduces a breakthrough in digital holography by applying an advanced instance segmentation approach, utilizing a modified Mask DINO architecture to significantly improve particle detection and segmentation. Facing the challenge of limited data, we leverage synthetic data generated by our proprietary FakeHolo system, enabling the use of data-intensive transformer-based models. Our approach yields substantial improvements in accuracy and efficiency, as demonstrated by enhanced recall, precision, and F1 scores, along with reduced sizing error across various datasets. This work not only advances the field of digital holography but also offers significant potential for applications requiring precise imaging and analysis, marking a notable step forward for both research and practical applications.