Cross-Family U-Net Landmark Heatmap Regression For Butterfly Wings -- Context and Generalization by Grouping

Abstract

This thesis presents a novel approach to automated landmark detection on butterfly wings using deep learning techniques, addressing the challenge of cross-family generalization in morphometric analysis. Accurate landmark detection is essential for studying morphological variations in butterflies, but traditional manual annotation is time-consuming and impedes large-scale research. We propose several improvements to U-Net-based landmark detection through anatomically-informed grouping strategies and enhanced preprocessing. By integrating YOLOv8 for butterfly detection and cropping before landmark prediction, we significantly improve input quality. We explore multiple model configurations, comparing single-landmark versus multi-landmark channel approaches, different loss functions (MSE and FBCE), and varying input resolutions (256×256 and 512×512). Our experiments on specimens from the Papilionidae family demonstrate that anatomically-guided multi-landmark grouping achieves convergence four times faster than conventional approaches while maintaining comparable accuracy. Higher resolution models (512×512) show substantially improved precision when evaluated at original image scale. Most importantly, our cross-family generalization experiments reveal that models trained on the combined Papilionidae and Morpho datasets with strategic landmark grouping and index matching successfully adapt to varying numbers of landmarks across butterfly families. These findings advance morphometric analysis capabilities in Lepidopterology and demonstrate the value of incorporating domain knowledge into neural network architecture design for biological feature detection tasks.