To Bee or Not to Bee : Pollinator Classification with ML & Optimization

IG.2412 — Machine Learning (2024/2025) Master's Project (Engineering cycle, ISEP Paris) — Machine Learning and Optimization.

Overview

We tackle pollinator insect classification from images and binary masks, addressing two tasks:

• Binary: Bee vs. Bumblebee
• Multiclass: Bee, Bumblebee, Others (Hover fly, Wasp, Butterfly grouped)

The pipeline combines classical machine learning on engineered features and a deep learning approach based on a pre-trained vision backbone. The goal is to compare approaches, visualize separability, and quantify performance.

Data & Preprocessing

• Images with pixel masks and class labels (IDs 1–347).
• Largest connected component kept from the mask; image cropped to its bounding box to remove background.
• All feature extraction and deep models operate on the masked regions.

Feature Engineering

From masks and masked RGB/grayscale images we derive:

• Shape: area, perimeter, circularity, eccentricity, aspect ratio, solidity.
• Texture (GLCM): homogeneity, energy (computed at angles 0, 45, 90, 135 degrees).
• Color: per-channel min, max, mean, median, standard deviation; "bug ratio" (sum RGB / mask size).
• Geometric:
  • Largest inscribed circle inside the mask using the Euclidean Distance Transform for initialization and Nelder–Mead to optimize center and radius under constraints.
  • Best symmetry axis via 1D rotation search around the circle center: rotate, mirror around the vertical through the rotated center, and minimize pixel-wise asymmetry with Nelder–Mead.

Classical ML

• Feature selection: SFS, RFE, and SelectKBest (ANOVA and Mutual Information). SFS found a compact subset (≈6–7 features) with strong validation accuracy.
• Models evaluated: k-NN, Decision Tree, SVM, Random Forest, XGBoost, Logistic Regression.
• Highlights:
  • Binary: 3-NN with SFS-selected features reached 93% validation accuracy.
  • Multiclass: XGBoost with all features reached 88% validation accuracy.

Unsupervised Clustering

K-Means, Agglomerative, Spectral clustering were explored. Silhouette scores remained low (~0.17–0.26), reflecting overlapping class structure and noisy separability.

Deep Learning

• Backbone: Pre-trained DINOv2 used as a frozen feature extractor.
• Head: Two linear layers with ReLU and dropout on the CLS token output.
• Training: 15 epochs, AdamW, cosine annealing LR, standard augmentations; masked images resized to 224×224; class imbalance mitigated by augmentation and oversampling.
• Results: 95% accuracy (binary) and 94% (multiclass) on validation; the DL approach outperformed classical baselines while remaining lightweight to train.

Takeaways

• Simple geometric and texture features already offer competitive baselines (especially for binary).
• DINOv2 features provide a strong boost for both tasks with minimal fine-tuning.
• Class overlap and label imbalance challenge unsupervised clustering and linear separability.

Acknowledgments

Course staff and public pretrained backbones (DINOv2).