Yang Yi, Li WenJie, Liu RuiZe, Wu ChengZhe, Ren Jing, Shi YiShi, Ge SiQin
Key Laboratory of Zoological Systematics and Evolution, Institute of Zoology, Chinese Academy of Sciences, 1 Beichen West Road, Chaoyang District, Beijing, 100101, China.
Department of Scientific Research, Beijing Planetarium, Xizhimenwai Street, Beijing, 100044, China.
Sci Data. 2025 Apr 23;12(1):680. doi: 10.1038/s41597-025-05010-y.
The utilization of datasets from beetle hindwings is prevalent in research of morphology and evolution of beetles, serving as a valuable tool for comprehending the evolutionary processes and functional adaptations under specific environmental conditions. However, the collection of hindwing images of beetles poses several challenges, including limited sample availability, complex sample preparation procedures, and restricted public accessibility. Recently, a machine learning technique called Stable Diffusion has been developed to statistically generate diverse images using a pretrained model with prompts. In this study, we introduce an approach utilizing Stable diffusion and ControlNet for the generation of beetle hindwing images, along with the corresponding results obtained from its application to a diverse set of 200 leaf beetle hindwings. To demonstrate the fidelity of the synthetic hindwing images, we conducted a comprarative analysis of three key metrics: Structural Similarity Index (SSIM), Inception Score (IS), and Fréchet Inception Distance (FID), which are crucial for evaluating image fidelity. The results demonstrated a strong alignment between the actual data and the synthetic images, confirming their high fidelity. This novel library of leaf beetle hindwings not only offers morphological image for utilization in machine learning, but also showcases the extensive applicability of the proposed methodology.
甲虫后翅数据集在甲虫形态学和进化研究中被广泛应用，是理解特定环境条件下进化过程和功能适应的宝贵工具。然而，甲虫后翅图像的收集面临诸多挑战，包括样本可用性有限、样本制备程序复杂以及公众获取受限。最近，一种名为稳定扩散的机器学习技术被开发出来，用于使用带有提示的预训练模型统计生成多样化的图像。在本研究中，我们介绍了一种利用稳定扩散和控制网络生成甲虫后翅图像的方法，以及将其应用于200种不同叶甲后翅所获得的相应结果。为了证明合成后翅图像的逼真度，我们对三个关键指标进行了比较分析：结构相似性指数（SSIM）、 inception得分（IS）和弗雷歇 inception距离（FID），这些指标对于评估图像逼真度至关重要。结果表明实际数据与合成图像高度一致，证实了它们的高逼真度。这个新的叶甲后翅库不仅为机器学习提供了形态图像，还展示了所提出方法的广泛适用性。
