Real-Time 3-D Particle Localization via Deep Learning-Based Image Deblurring for Enhanced Measurement Resolution

High-precision 3-D particle localization under limited depth of field remains a critical challenge in microscopic imaging, particularly in dynamic measurement scenarios. In this study, we propose a deep learning-based framework aimed at enhancing spatial resolution for particle tracking from defocused images. The network is built upon an improved DeblurGAN-v2 architecture, which incorporates an attention mechanism and custom loss functions specifically optimized for high-accuracy 3-D localization. The model is then trained and validated on a hybrid dataset comprising both experimentally acquired and synthetically simulated particle images, ensuring robust image restoration performance across a wide range of defocus conditions. The proposed framework is integrated into a real-time localization system, supporting particle tracking at approximately 43 Hz. Experimental validation, including both static particle localization and dynamic 3-D trajectory tracking, demonstrates over a 40% improvement in both lateral and axial resolution. This strategy provides a practical and effective solution for accurate 3-D particle localization overextended axial ranges, with strong potential for applications in micro/nanoscale measurement and control systems.