Deep Learning Methodology for Wood Species Identification using High-Resolution Macroscopic Images and Patch-Voting

A deep learning framework for timber classification using macroscopic images with patch-based ensemble voting.

Key Features

• 🔬 High-resolution macroscopic image processing
• 🗳️ Patch-based majority voting system

Installation

# Clone repository
git clone https://github.com/ari-dasci/S-GoIMAI.git
cd S-GoIMAI

Dataset

The GOIMAI Phase-I dataset of high-resolution macroscopic wood images is available on Zenodo:

Download the dataset from:
https://zenodo.org/records/11092208

This expanded version contains:

• 8.7 GB of high-resolution macroscopic images
• 37 wood species with multiple samples each
• 24x magnification captures using specialized equipment

After downloading, unzip the dataset and use the directory structure as:

raw_dataset/
├── species_1/
│   ├── image1.jpg
│   └── image2.jpg
└── species_2/
    ├── image1.jpg
    └── image2.jpg

Usage

1. Data Augmentation

Generate augmented dataset from original images:

python scripts/create_dataset.py \
  -i path/to/original_dataset \
  -o path/to/augmented_output \
  -n 10 \                 # 10 augmented versions per image
  -t 8 \                  # Use 8 CPU threads
  -v                       # Verbose output

2. Model Training

Train a classification model with k-fold cross-validation:

python scripts/train.py \
  --input path/to/augmented_dataset \
  --original path/to/original_dataset \
  --model b0 \            # EfficientNetV2-B0 architecture
  --size 300 300 \        # Input image dimensions
  --output path/to/models \
  --epochs 100 \
  --batch 32 \
  --k_fold 5 \            # 5-fold cross-validation
  --loss sparse_categorical_crossentropy

Key Arguments:

Argument	Description	Example
--input	Path to augmented dataset	path/augmented_data
--original	Path to original unaugmented dataset	path/original_data
--model	Model architecture (b0,b1,b2,b3,ir2)	b2
--size	Input dimensions (width height)	224 224
--k_fold	Cross-validation folds	5
--output	Model output directory	path/models

3. Output Structure

After training, models are saved with metadata:

models/
└── 2024-06-25_14:30:00-b0-sparse_categorical_crossentropy-100epochs/
    ├── fold_0-b0.keras
    ├── fold_0-b0-history.csv
    └── metadata.json

Features

• 🌀 Automated data augmentation pipeline
• 🔢 Stratified k-fold cross-validation
• 📈 Cosine learning rate decay
• 🧠 Support for EfficientNetV2 and InceptionResNetV2
• 📊 Patch-based evaluation with majority voting

Requirements

• Python 3.8+
• TensorFlow 2.10+
• Pillow
• scikit-learn
• pandas