A survey of neural network segmentation and validation on plant specimen images of Korean Violaceae

Abstract

Accurate plant identification is crucial for biodiversity research, yet manual classification remains time-consuming and requires specialized expertise. To overcome these challenges automated identification technologies are increasingly being developed. A key step in this process is the precise segmentation of plant materials from plant specimen images; however, existing approaches often struggle to separate plant material from non-plant components such as labels, barcodes, stamps, and rulers. To address this problem, we propose integrating Multi Receptive Field (MRF) blocks into a U-Net framework, enabling robust multi-scale feature extraction from plant bodies of varying sizes in digitized specimens. We conduct extensive experiments on a dataset of 14,939 plant specimen images from 36 species of Viola (Violaceae), comparing the performance of eleven segmentation models, including ten state-of-the-art methods and our approach. The proposed model achieved superior performance with a mean Intersection over Union of 0.8531, Dice coefficient of 0.9123, and pixel accuracy of 0.9920. Through ablation studies, we established that incorporating six different kernel sizes in the MRF block yields optimal segmentation results. By effectively addressing the complexities inherent in herbarium images-such as varying plant scales and the presence of non-plant elements–our model establishes a strong foundation for automated plant identification. This study advances digital herbarium analysis and highlights the potential of specialized neural network architectures in botanical research.