Title: DefocusTrackerAI: A Universal Framework for the Automated Identification of Defocused Particle Imagery

Abstract

This study presents DefocusTrackerAI, an advanced deep-learning framework designed for the automatic localization and detection of defocused particle images across diverse optical configurations. Serving as the successor to the open-source DefocusTracker project, this new system maintains high recall rates and minimal uncertainty. To optimize performance, we evaluated two prominent object detection architectures, YOLOv9 and Faster R-CNN, using a comprehensive synthetic dataset featuring astigmatic and non-astigmatic defocused particles of various sizes.

Comparative analysis revealed that YOLOv9 surpasses Faster R-CNN, delivering superior recall and reduced uncertainty, especially under conditions of high particle image density. Furthermore, YOLOv9 demonstrated improved spatial resolution, recording uncertainty levels between 0.1 and 0.4 pixels for particle densities ($N_s$) up to 0.5. This performance exceeds that of current state-of-the-art algorithms.

We validated the models' robustness by detecting both astigmatic and non-astigmatic defocused particles across multiple optical setups with varying illumination. The framework’s versatility was further confirmed through successful application in real-world Defocused Particle Tracking (DPT) experiments, encompassing both fluorescence and shadowgraph data. These results indicate that the tool extends beyond traditional DPT uses, enabling the effective tracking of sprays and droplets.

A pre-trained, ready-to-deploy version of DefocusTrackerAI, built on the YOLOv9 architecture, is publicly accessible at https://gitlab.com/goncalo.coutinho/defocustrackerAI-main/-/tree/7e0f11f649ebad50e20dca5b9545f26ca303ebe0. This resource allows for the high-accuracy automatic detection of defocused particle images of any type. When integrated with an appropriate depth-position calibration method, it serves as a powerful initial step for three-dimensional defocusing particle tracking.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC