Source-linked AI summary
Deep-Plant: Plant Identification with convolutional neural networks
Sue Han Lee, Chee Seng Chan, Paul Wilkin, Paolo Remagnino
TL;DR
Plant identification is difficult because leaves from different species can have similar appearances and hand-crafted features may be dataset-dependent. This paper fine-tunes CNNs on 44 plant species and uses DN visualisation to inspect the learned features. The model reaches 99.6% accuracy, with different-order venations identified as important features, while environmental leaf damage contributes to misclassification.
Problem
Plant identification remains challenging because species can have similar leaf shape and colour, while hand-crafted feature performance is task- or dataset-dependent.
Method
The paper fine-tunes CNNs for 44-class leaf classification and applies DN visualisation to examine the learned feature representations.
Results
99.6% classification accuracy is reported for D2, and different-order venations are identified as robust features for plant identification.
Takeaways & Limitations
CNN-learned features provide better leaf-image representations than hand-crafted features, while DN visualisation reveals venation structure as important for distinguishing plant species.
Takeaways & Limitations
Environmental conditions such as wrinkled surfaces and insect damage contribute to misclassified leaves in D2.
Abstract
from arXiv · showhide
This paper studies convolutional neural networks (CNN) to learn unsupervised feature representations for 44 different plant species, collected at the Royal Botanic Gardens, Kew, England. To gain intuition on the chosen features from the CNN model (opposed to a 'black box' solution), a visualisation technique based on the deconvolutional networks (DN) is utilized. It is found that venations of different order have been chosen to uniquely represent each of the plant species. Experimental results using these CNN features with different classifiers show consistency and superiority compared to the state-of-the art solutions which rely on hand-crafted features.
1. INTRODUCTION
Plant identification remains difficult because species often have similar leaf shape and colour, while hand-crafted features can be task- or dataset-dependent. The paper proposes CNN feature learning with DN visualisation to classify 44 species and understand which leaf features the model uses.
- Plant identification remains challenging because different species can share similar leaf shape and colour.
- Existing approaches use hand-crafted shape, geometric, colour, texture, and venation features, whose performance depends on task- or dataset-specific choices.
- The paper uses CNNs to automatically learn leaf representations for classifying 44 plant species.
- Deconvolutional networks visualise learned CNN features so the model is not treated as a black box.
- The deep-learning framework combines bottom-up CNN classification with top-down visual analysis of plant-identification features.
- The study introduces the fully annotated 44-class MalayaKew Leaf Dataset collected at the Royal Botanic Gardens, Kew, England.
2. PROPOSED APPROACH
The approach reuses and fine-tunes a pre-trained CNN for 44-class leaf identification, then applies a deconvolutional-network visualization strategy to inspect its activated regions and failures.
- 2. PROPOSED APPROACH: The framework uses a pre-trained CNN and fine-tunes it for plant identification rather than training a new architecture.The network was pre-trained on ILSVRC2012 and reused because generic CNN features can transfer and the plant training set is smaller.
- 2. PROPOSED APPROACH: Fine-tuning replaces the original 1000-way output with 44 neurons for the Kew leaf dataset.The dataset contains 44 classes collected at the Royal Botanic Gardens, Kew, England.
- 2. PROPOSED APPROACH: The visualization includes failure-analysis figures for the CNN model on datasets D1 and D2.These figures are presented as part of the strategy for understanding how and why the CNN works or fails.
- 2. PROPOSED APPROACH: The CNN learns filters through back propagation, whose activations extract features intended to represent each input leaf image.Filter visualization is used to inspect feature transformations and the model’s internal operation rather than treating it as a black box.
- 2. PROPOSED APPROACH: V1 uses a deconvolutional-network strategy to reconstruct images from the highest activations across a layer’s feature maps.Only the first S largest absolute activation values are retained, while the rest are set to zero before projection into pixel space.
3. EXPERIMENTAL RESULTS
CNN features achieved strong plant-identification accuracy on the 44-class MK Leaf Dataset, while DN visualisation showed that excluding leaf shape shifted attention toward multi-order venation.
- Performance: 98.1% accuracy from CNN-learned features outperformed state-of-the-art methods using hand-crafted features across different classifiers.The comparison was conducted on the MK Leaf Dataset.
- D1 failure analysis: DN visualisation of D1 showed strong activation on leaf shape, which the authors deduced was not a good identification feature.The observation came from reconstructed images using the V1 strategy.
- D2 feature analysis: For D2, activations covered primary and secondary venation and departures between venation orders.The authors deduced that venation of different orders provides more robust features for plant identification.
- D2 failure analysis: D2 misclassifications were concentrated in Classes 9 and 2, with environmental damage and wrinkled leaf surfaces identified as contributing conditions.Class 9 had 18 misclassified patches and Class 2 had 13.
4. CONCLUSION
The paper concludes that CNNs can learn discriminative leaf features more effectively than hand-crafted representations for plant identification. DN analysis identifies multi-order venation as an important feature, with 99.6% performance on the studied dataset.
- CNN-learned features provided better leaf-image representations than hand-crafted features for plant identification.
- 99.6% performance was achieved while identifying venation structure as an important feature for distinguishing plant species.
- DN visualisation verified the venation finding by analysing the network’s internal operation and behaviour.