This work introduces an innovative intra-class morphing data augmentation method designed to improve insect classification in low-data scenarios. Leveraging As-Rigid-As- Possible (ARAP) deformation and landmark detection, the approach combines controlled geometric warping with realistic texture interpolation to generate synthetic images that remain faithful to species-specific morphological characteristics. The method was rigorously validated, both qualitatively and quantitatively, on the Leeds Butterfly Dataset and the Butterfly Moths Image Classification (100 Species) Dataset. When integrated into an InceptionV3 network pre-trained on ImageNet, the proposed augmentation leads to a significant improvement in classification performance, consistently outperforming traditional techniques such as Mixup, Cutout, and Rotation across key metrics, including precision, recall, accuracy, and F1-score.
ELBEY et al. (Thu,) studied this question.