Related Experiment Video
Updated: Aug 25, 2025
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
9.3K
Plant recognition by AI: Deep neural nets, transformers, and kNN in deep embeddings
Lukáš Picek1, Milan Šulc2, Yash Patel2
1Department of Cybernetics, Faculty of Applied Sciences, University of West Bohemia, Pilsen, Czechia.
Frontiers in Plant Science
|October 14, 2022
Summary
A new retrieval-based machine learning method outperforms current Convolutional Neural Networks (CNN) and Vision Transformers (ViT) for plant species recognition. This approach offers superior accuracy on large datasets, advancing botanical identification technology.
Area of Science:
- Computer Science
- Botany
- Machine Learning
Background:
- Automatic plant species recognition is crucial for biodiversity monitoring and ecological research.
- Existing machine learning methods, including Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), have shown promise but face challenges with large-scale, diverse datasets.
- Benchmarking these state-of-the-art approaches is essential for identifying areas for improvement.
Purpose of the Study:
- To review and benchmark current machine learning methods for image-based plant species recognition.
- To propose and evaluate a novel retrieval-based method for plant recognition using nearest neighbor classification in a deep embedding space.
- To assess the impact of various performance-enhancing techniques on recognition accuracy.
Main Methods:
- A novel retrieval-based method utilizing nearest neighbor classification in a deep embedding space, trained with the Recall@k surrogate loss.
- Benchmarking of state-of-the-art Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) against the proposed method.
- Evaluation of performance-enhancing techniques such as class prior adaptation, image augmentations, learning rate scheduling, and loss functions.
Main Results:
- The proposed retrieval-based method achieved superior performance across all evaluated datasets (PlantCLEF 2017, ExpertLifeCLEF 2018, iNaturalist 2018), outperforming CNN and ViT classifiers.
- State-of-the-art Vision Transformer (ViT-Large/16) achieved 91.15% accuracy on PlantCLEF 2017 and 83.54% on ExpertLifeCLEF 2018.
- The retrieval approach demonstrated accuracy margins of 0.28%, 4.13%, and 10.25% over other methods on ExpertLifeCLEF 2018, PlantCLEF 2017, and iNat2018-Plantae, respectively.
Conclusions:
- The novel retrieval-based method represents a significant advancement in automatic plant species recognition.
- This approach offers improved accuracy and robustness compared to existing classification methods, particularly on large and complex datasets.
- Further research can explore optimizing retrieval-based techniques for enhanced botanical identification systems.
