Exploring Image Classification in Herbarium Datasets: Visualizing the Effects of Image Perturbations
Open AccessThis research addresses the challenge of understanding model accuracy and generalization in fine-grained datasets, specifically focusing on herbarium collections from 2019 and 2022. The study developed a novel process to identify the most effective model for classification tasks, utilizing visualizations to pinpoint key features that models learn from data. A comparative analysis of several models was conducted to establish consistency in feature recognition, aligning with known morphological characteristics. The research further explored the robustness of these models by simulating outliers through abnormal normalization techniques.The study's innovative approach also included an analysis of model performance under various data perturbations, such as Adversarial Noise, Gaussian Blur, and Max Filter. These experiments aimed to understand the impact of such perturbations on model reliability. The findings indicated that specific preprocessing choices, such as resolution adjustments, data augmentation, and color processing, significantly affect validation accuracy. The use of visualization tools revealed that model accuracy varied with the complexity of plant structures and the presence of damaged specimens within the datasets. This research contributes to the field by providing insights into the effectiveness of image classification models on fine-grained herbarium datasets and highlighting the importance of selecting appropriate preprocessing techniques to enhance model accuracy and robustness.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.