DIATLAS - The French freshwater DIatom ATLAS image dataset

This dataset contains photonic (optical) microscopy images of diatoms extracted from French regional atlases, documenting specimens collected from freshwater streams. These images were acquired using either Bright Field (BF) or Differential Interference Contrast (DIC) techniques. Each image is associated with a taxonomic classification. This dataset was used to train two Convolutional Neural Network (CNN) models: A detection model, based on the ai4oshub/ai4os-yolov8-torch module, designed to predict oriented bounding boxes around diatoms. It is built on a pre-trained YOLOv8, fine-tuned on manually labeled data. > Diatom detection with oriented bounding boxes A classification model, based on Ultralytics' YOLOv8-cls architecture, designed to classify diatom at the species level. > Diatom classification at the species level Image metadata are provided in the diatoms.csv, sources.csv, and taxonomic_code.csv files.