Yramklass/POL-ID

POL-ID: A deep learning pipeline that authenticates South African honey origins by automatically detecting and classifying pollen grains from microscope slides. # POL-ID: Deep Learning for Pollen Detection and Classification in the Authentication of South African Honey **POL-ID** is an end-to-end deep learning pipeline designed to automate the process of melissopalynology (the study of pollen in honey). By analyzing microscopic images of honey slides, this project detects and classifies pollen grains to determine the honey's floral and geographical origin, providing a powerful tool for quality control and authentication. The system leverages state-of-the-art computer vision models to provide fast, accurate, and reproducible results, significantly reducing the time and expertise required compared to traditional manual methods. --- ## Key Features * **High-Performance Pollen Detection:** Utilizes object detection models like **YOLOv8** and **DETR** to accurately locate and bound individual pollen grains on a microscopic slide. * **Advanced Pollen Classification:** Employs a novel **parallel fusion architecture** combining a **Swin Transformer** and a **ConvNeXt** model for robust and highly accurate species classification. * **Smart Clustering of Unknowns:** Integrates **UMAP** for dimensionality reduction and **HDBSCAN** to intelligently group low-confidence pollen grains, helping to identify rare types or highlight potential new species not present in the training data. * **Comprehensive Reporting:** Generates a detailed final report including total pollen counts, species ratios (percentages), and a final honey classification (e.g., monofloral, multifloral). --- ## Data Preprocessing The raw data used for this project required significant preprocessing to ensure model training was effective and evaluation was fair. Key challenges included: * **Inconsistent Labeling:** The raw detection data had multiple class labels for pollen, which were standardized to a single class (`0`). The classification labels also had some inconsistencies (e.g. misspellings, incorrect/inconsistent class labels). These were handled throug …