Few-Shot Learning for Automated Sexual Dimorphism Classification in Scorpions Using Multi-View Images

Morphological differences between sexes represent a fundamental topic in biology and are observed in scorpions as in other organisms, defined by variations in body part proportions, sizes, and shapes between males and females. Sex determination in scorpions presents challenges for biologists yet provides critical insights into evolutionary processes, ecological adaptations, and reproductive strategies. Field researchers encounter difficulties distinguishing sexes, particularly given the venomous nature of some species. This study aimed to classify scorpion sexes using dorsal and ventral images through a few-shot learning approach with prototypical networks for sex classification with limited labeled data. The dataset comprised 99 Aegeaeobuthus gibbosus scorpions (50 females, 49 males), with both dorsal and ventral images obtained for each specimen. DenseNet121 and ResNeXt-50 architectures, pre-trained on ImageNet, served as base networks and were evaluated at 1-shot, 5-shot, and 10-shot settings. Experimental results indicated that DenseNet121 achieved 84.94% accuracy at the 10-shot setting, while ResNeXt-50 achieved 84.14%. Both architectures demonstrated consistent accuracy improvements as support samples increased, confirming the few-shot learning framework's effectiveness. Fold-based confusion matrix analyses showed that DenseNet121 exhibited more stable performance with challenging samples, whereas ResNeXt-50 was more sensitive to individual morphological variation but achieved comparable accuracy with sufficient support samples. These findings demonstrate that few-shot learning provides a viable solution for scorpion sex classification with limited data. High accuracy can be achieved using a small number of labeled specimens, offering a practical approach for researchers studying rare or endangered scorpion species.