Few-Shot Learning for Microscopic Wood Species Classification

Overview

This project implements a Few-Shot Learning (FSL) pipeline for microscopic wood species classification, designed to mimic the low-sample, high-class-imbalance conditions found in archaeological charcoal identification tasks.

The work is inspired by the AI-WOOD project goals and demonstrates how meta-learning techniques can be used to train a classifier that generalizes to unseen wood species with only a few labeled examples.

The complete implementation is based on PyTorch and supports:

• Configurable Few-Shot parameters (N-way, K-shot, Q-query)
• Episodic training setup
• OneCycle learning rate policy for faster convergence
• Detailed metrics tracking (accuracy, precision, F1-score)
• Visualization of confusion matrices for interpretability

---

Dataset

We use the Forest Species Database (FSD) — a publicly available dataset containing microscopic images of various wood species.

Dataset details:

• Total images: 2,240
• Number of species: 112
• Images per species: 20
• Image type: Microscopic cross-sections

This dataset closely resembles the microscopy context described for ancient charcoal classification.

Example Images:

---

Repository Structure

.
├── src/                     # Source code for dataloader, model, and loss functions
│   ├── dataloader.py         # Few-shot episodic data loading
│   ├── model.py              # ResNet-based embedding generator + dense 
│   └── loss.py               # Prototypical loss / metric learning losses
│               # Utility functions for metrics and visualization
│
├── train.ipynb               # Main training notebook
├── helper.py                 # helper functions to load blocks
├── install.sh                # envirement installer 
├── README.md                 # Project documentation (this file)

---

Installation

1. Clone the repository

git clone https://github.com/Chaouki-AI/FSL-MicroWood.git
cd FSL-MicroWood

2. Create a virtual environment

chmod +x install.sh
./install.sh

---

Training

Run the main training notebook

jupyter notebook train.ipynb

You can configure the training parameters (N-way, K-shot, Q-query, image size, learning rate, etc.) via the Args class in train.ipynb.

---

Model Architecture

The model uses:

• Backbone: ResNet (pretrained on ImageNet)
• Embedding Layer: Dense layer projecting to 256-dimensional feature space
• Loss: Prototypical loss for few-shot classification
• Optimizer: AdamW
• Learning Policy: OneCycle Learning Rate Scheduler

---

Results

For a 5-shot, 15-query, 6-way configuration:

• Accuracy: 98%
• Precision: 95%

Training Curves:

Confusion Matrix:

---

Key Features

• Episodic Training: Learns how to adapt quickly to new species with minimal examples.
• Parameter Flexibility: All Few-Shot hyperparameters are adjustable.
• Interpretability: Confusion matrix visualization to assist domain experts.
• Extendability: Easily adaptable to other microscopic image classification tasks.

---

How to Adapt to Charcoal Classification

The pipeline is dataset-agnostic — simply replace the dataset in src/dataloader.py with charcoal microscopic imagery and adjust preprocessing as needed. The model and episodic training logic remain unchanged.

---

License

This project is released under the MIT License.