An Improved Method for Crystal Morphology Extraction Using In Situ Imaging

This study proposes an enhanced Improved Refined Rotation RetinaNet (iR 3 Det) to address overlapping crystal detection challenges in crystallization processes. By integrating multi‐scale Retinex image enhancement for improved visibility and fast neural architecture search for object detection (NASFCOS)‐Feature Pyramid Network (FPN) into the neck network for refined feature representation, the method reduces missed detections while resolving crystal overlaps. Validated on in situ images of β‐ l ‐glutamic acid (β‐LGA) during seeded batch‐cooling crystallization, iR 3 Det demonstrates superior accuracy in extracting morphological parameters from adherent crystals compared to original R 3 Det and traditional segmentation methods. The algorithm effectively balances detection precision for both isolated and overlapping crystals, enabling reliable crystal size distribution analysis critical for crystallization process monitoring and control. This approach provides a robust solution for real‐time crystal morphology characterization in dynamic industrial crystallization environments.