Radiolarian Microfossil Dataset for DSG-MSAF: 32-class SEM images

This dataset contains scanning electron microscopy (SEM) images of radiolarian microfossils used in the paper "Decaying Synthetic Guidance and Multi‑Scale Adaptive Fusion for Long‑Tailed Microfossil Classification" (submitted to Engineering Applications of Artificial Intelligence). The dataset comprises 32 taxonomic classes of radiolarians, with a long‑tailed class distribution typical of real‑world paleontological collections. After preprocessing, each class was capped at 1,000 images, resulting in a total of 20,804 images (7,569 real SEM images + 13,235 high‑fidelity synthetic images generated via FLUX.1 with Canny conditioning). The synthetic images were curated using a CLIP‑based dual evaluation to ensure morphological consistency. The data is split into: Training set: 80% (stratified, preserving class distribution) Validation set: 10% Test set: 10% All images are resized to 384 × 384 pixels and saved in PNG format. File names follow the pattern [class]_[unique_id].png. This dataset is intended for research on long‑tailed classification, fine‑grained visual recognition, and generative data augmentation in scientific imaging. If you use this dataset, please cite the accompanying paper.