PaperlayerView exact revision on arXiv

Source-linked AI summary

GANet: A Lightweight Group Attention Network for Plant Disease Identification on Embedded Systems

Anubhab Tripathi, Bibek Das, Asfak Ali, Sheli Shinha Chaudhuri

arXiv:2610.30225v1cs.LG

TL;DR

Plant-disease monitoring calls for timely identification, while conventional monitoring techniques can prevent early intervention by local farmers. The paper proposes a lightweight group-equivariant classification model that combines symmetry handling with feature fusion and attention. On the testing dataset, it reports 99.75% accuracy and a loss of 0.009.

  • Problem

    Conventional disease-monitoring techniques can prevent early intervention by local farmers, despite the need for timely plant-disease identification.

    Source: §I · p. 1

  • Method

    The proposed model combines group-equivariant convolutions with feature fusion and attention, using shared filters to reduce parameters.

    Source: B. Proposed Model · p. 3, 1) Feature Fusion Block (FFB) · p. 3, 2) Group-Attention Block (GAB) · p. 4

  • Results

    99.75% accuracy and 0.009 loss were reported on the testing dataset.

    Source: A. Experimental Result · p. 4

  • Takeaways & Limitations

    The authors target an image classifier that is easy to implement and has fewer parameters than existing state-of-the-art techniques while maintaining accuracy.

    Source: §V · p. 7

    Integration with real-time image processing remains future work intended to address latency and bandwidth.

    Source: §V · p. 7

Abstract

from arXiv · show

Global warming and climate change are major catalysts for plant diseases, significantly affecting food production worldwide. Addressing this issue requires continuous monitoring and early detection of plant diseases. While recent advancements in deep learning offer potential solutions, most models demand high processing power and struggle to perform efficiently on low-end electronic hardware. Furthermore, many existing works focus on specific crop diseases, limiting their broader applicability. This paper presents a lightweight Group Equivariance Attention Network (GANet) designed for the prediction and monitoring of 38 types of plant diseases. GANet introduces a novel Group Attention block and Group-Wise Feature Fusion block, making it both sparse and efficient. With only 0.001M parameters, the model achieves an impressive 99.99% accuracy on the PlantVillage dataset. Additionally, GANet is optimized for deployment on edge devices such as the Raspberry Pi 4b and Jetson Nano, where it reaches 5 FPS and 11 FPS, respectively, demonstrating its potential for real-time plant disease monitoring on low-power devices.

I. INTRODUCTION

Plant-disease identification matters for timely intervention, while conventional monitoring techniques can prevent early action by local farmers [3]. The paper proposes a lightweight group-equivariant network, motivated by the potential to handle image symmetries with fewer parameters and lower computational cost.

I. INTRODUCTION

  • The proposed lightweight Group Equivariance attention network targets plant-disease identification with fewer parameters, lower computational costs, and simpler implementation.The motivation includes exploiting transformations such as rotation and symmetry rather than merely retaining image information.Source: §I · p. 1

II. LITERATURE SURVEY

  • Prior plant-disease work spans deep-learning models, edge-device deployment, and reviews of methods whose reported accuracies range from 80% to 99%.The cited review range depends on the particular algorithm and dataset.Source: §II · p. 1

A. Foundation of Group Convolution

  • Group-equivariant convolutions extend convolutional symmetry handling beyond translations to transformations such as rotations and reflections, [22].The paper gives D4, with rotations by multiples of 90° and reflections, as an example.Source: A. Foundation of Group Convolution · p. 2
  • A G-equivariant convolution maps a transformed input to a predictably transformed feature map.The equation states Φ(g · x) = g · Φ(x) for all g ∈ G.Source: A. Foundation of Group Convolution · p. 2

1) p4convp4 Convolution:

The p4convp4 layer is designed to be equivariant under D4, the square-symmetry group of rotations and reflections. Its shared-filter formulation applies across transformations in that group.

1) p4convp4 Convolution:

  • D4 comprises square symmetries, including rotations by multiples of 90° and reflections.The presentation identifies a as a counterclockwise 90° rotation, b as a reflection, and e as the identity.Source: 1) p4convp4 Convolution · p. 2
  • In p4convp4, filters are shared across all D4 transformations, and the output feature map is computed from the transformed kernel and input.The passage identifies k as the filter kernel, g−1h as its transformation, and h as a group element applied to the input.Source: 1) p4convp4 Convolution · p. 2
  • A single filter adapts to all group transformations, reducing parameter count and improving generalization.This is the stated effect of the formulation.Source: 1) p4convp4 Convolution · p. 2

2) p4convz2 Convolution:

The p4convz2 layer extends p4convp4 with an additional Z2 symmetry group for binary reflections. The paper describes the combined group as supporting richer symmetry handling and generalization across orientations and reflections.

2) p4convz2 Convolution:

  • p4convz2 adds Z2, representing binary reflections, to p4convp4's symmetry handling.The combined group is written as Z2 × D4.Source: 2) p4convz2 Convolution · p. 2
  • The p4convz2 operation uses a transformed filter kernel that accounts for both rotation and reflection symmetries.The passage also describes the input feature map as transformed under the combined group actions.Source: 2) p4convz2 Convolution · p. 2
  • Group-equivariant convolutions use one filter for patterns across transformations, reducing parameters and mitigating overfitting, particularly with limited training samples.The paper also states that handling rotations and reflections inherently can eliminate the need for extensive data augmentation.Source: 2) p4convz2 Convolution · p. 3

B. Proposed Model

The proposed plant-disease classifier combines group-equivariant convolutions with feature fusion and attention. Its Feature Fusion Block sums outputs from P4ConvZ2 and P4ConvP4 layers to aggregate features while preserving equivariance.

B. Proposed Model

  • The model combines group-equivariant convolutions, a Feature Fusion Block, and an Attention Block for plant disease classification.It uses P4ConvZ2 and P4ConvP4 convolutions over G = D4 × Z2 to handle rotations and reflections.Source: B. Proposed Model · p. 3

1) Feature Fusion Block (FFB):

  • The Feature Fusion Block aggregates P4ConvZ2 and P4ConvP4 outputs by element-wise summation.The resulting feature map combines information from multiple transformations.Source: 1) Feature Fusion Block (FFB) · p. 3
  • The sum of two group-equivariant mappings remains group-equivariant under the same group action.The proof sketch attributes this preservation to addition’s linearity and the equivariance of both mappings.Source: 1) Feature Fusion Block (FFB) · p. 3

2) Group-Attention Block (GAB):

The Group-Attention Block refines feature components through softmax-weighted self-attention and combines them by element-wise multiplication. It is presented as complementing the G-CNN features by emphasizing informative regions and reducing noise.

2) Group-Attention Block (GAB):

  • The Group-Attention Block applies softmax-weighted self-attention to four components of the input feature map.Its attention weights emphasize relevant regions within each component.Source: 2) Group-Attention Block (GAB) · pp. 3–4
  • The block combines its refined components through element-wise multiplication.The resulting output is Ya.Source: 2) Group-Attention Block (GAB) · p. 4
  • The GAB is described as emphasizing critical regions, reducing noise, and complementing invariant features from the G-CNN layers.The stated hypothesis assumes attention preserves the structure induced by the group.Source: 2) Group-Attention Block (GAB) · p. 4
  • The proposed model integrates G-CNN symmetry handling, feature augmentation, and attention-based refinement.The paper reports superior robustness and accuracy in complex agricultural datasets.Source: 2) Group-Attention Block (GAB) · p. 4

C. Dataset

PlantVillage is a publicly accessible leaf-image dataset used for plant disease detection and classification. It contains more than 54,000 images across 14 crop species and 38 healthy and diseased classes.

C. Dataset

  • PlantVillage contains over 54,000 annotated leaf images from 14 crop species, organized into 38 healthy and diseased classes.The passage describes the dataset as diverse and balanced.Source: C. Dataset · p. 4
  • Table I is titled “Experimental Parameter Configuration.”The supplied passage identifies the table’s topic but does not provide its parameter values.Source: Table I · p. 4

D. Training and Hyperparameter Tuning

The training pipeline uses standardized, augmented 224 × 224 images and Adam optimization with a learning-rate scheduler. Training runs for up to 100 epochs, with dropout, validation monitoring, checkpointing, and early stopping.

D. Training and Hyperparameter Tuning

  • 224 × 224 images are augmented with rotations up to 15 degrees, random horizontal flips, and normalization.The passage says these augmentations simulate diverse real-world conditions.Source: D. Training and Hyperparameter Tuning · p. 4
  • Training uses Adam with an initial learning rate of 0.001, ReduceLROnPlateau, CrossEntropyLoss, 100 epochs, and batch size 128.The scheduler reduces the learning rate when validation loss plateaus.Source: D. Training and Hyperparameter Tuning · p. 4
  • A dropout rate of 0.5 is applied in the fully connected layers to mitigate overfitting.Training and validation losses and accuracy are monitored; checkpoints use validation-loss improvements, and early stopping halts training if validation loss stagnates.Source: D. Training and Hyperparameter Tuning · p. 4

IV. RESULT ANALYSIS

The model reports 99.75% test accuracy and a 0.009 loss, with results compared against other studies on PlantVillage. The confusion-matrix analysis describes class-level correct and incorrect classifications and reports accuracy alongside precision, recall, and F1.

A. Experimental Result

B. Confusion Matrix

  • The normalized confusion matrix places actual classes in rows and predicted classes in columns, with diagonal entries showing correct classifications and off-diagonal entries showing errors.Normalization addresses class imbalance when assessing performance across classes.Source: B. Confusion Matrix · p. 5
  • 99.75% accuracy is reported alongside precision, recall, and F1 as additional performance metrics.Accuracy is defined using true positives, true negatives, false positives, and false negatives; precision, recall, and F1 describe positive predictions and their balance.Source: B. Confusion Matrix · p. 6

C. GradCAM

Grad-CAM visualizations show image regions that the model considers important for classification, using color to indicate their relative importance.

  • Grad-CAM uses class-score gradients with final convolutional feature maps to create an overlaid heatmap, where warm colors indicate greater importance and cool colors less impact.Source: C. GradCAM · p. 6

D. Precision vs Recall Curves

The precision-recall analysis describes threshold-dependent trade-offs and substantial differences among classes, while the ROC analysis uses a one-vs-rest strategy across all 38 classes. The reported curves include near-perfect PR values for some classes and lower, irregular performance for Corn Cercospora Gray leaf spot.

D. Precision vs Recall Curves

  • Class-specific precision-recall curves assess the precision-recall trade-off across decision thresholds, with area under each curve summarizing that class’s performance.Source: D. Precision vs Recall Curves · p. 6
  • PR values are 1.0 for Apple healthy, Blueberry healthy, and many other classes, while Corn (maize) Cercospora Gray leaf spot has more irregular curves and lower AUPRC values.Source: D. Precision vs Recall Curves · p. 6

E. Multiclass ROC and Area Under Curves [AUCs]

F. t-SNE Visualization

The t-SNE discussion explains how pairwise similarities are preserved in a low-dimensional embedding for visualization. The conclusion describes the model’s group-equivariant approach, its parameter-efficiency goal, and directions for future work.

F. t-SNE Visualization

  • Fig. 6 presents multiclass ROC curves and their AUCs for all classes.Source: Figure 6 · p. 7
  • t-SNE maps high-dimensional data to a low-dimensional embedding by optimizing preservation of pairwise similarities.The described optimization uses gradient descent to minimize divergence and can produce clusters and sub-clusters.Source: F. t-SNE Visualization · pp. 6–7

V. CONCLUSION AND FUTURE SCOPE

  • The conclusion describes applying group-equivariant convolutions to PlantVillage disease classification and exploiting symmetry and weight sharing to improve model efficacy.The stated aim is an image classifier that can be implemented easily and with few constraints.Source: §V · p. 7
  • The model’s design goal is to reduce parameters relative to existing state-of-the-art techniques while maintaining accuracy.The conclusion attributes better accuracy to the layer combination and architectural design.Source: §V · p. 7
  • Future work proposes adding leaf-feature components, integrating real-time image processing to address latency and bandwidth, and exploring one-shot and few-shot learning across crop species.Source: §V · p. 7
Loading 2610.30225v1…