Related Experiment Video
Updated: Jul 10, 2025
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
9.2K
Recognition of Wheat Leaf Diseases Using Lightweight Convolutional Neural Networks against Complex Backgrounds
Xiaojie Wen1,2, Minghao Zeng1,2, Jing Chen1,2
1Key Laboratory of the Pest Monitoring and Safety Control of Crops and Forests of the Xinjiang Uygur Autonomous Region, College of Agronomy, Xinjiang Agricultural University, Urumqi 830052, China.
Life (Basel, Switzerland)
|November 25, 2023
Summary
Optimizing convolutional neural networks (CNNs) for wheat disease detection is crucial. The MnasNet model achieved 98.65% accuracy using SGD + StepLR training and a 0.001 learning rate, making it ideal for mobile disease identification.
Area of Science:
- Agricultural Science
- Computer Vision
- Machine Learning
Background:
- Wheat diseases pose a significant threat to global crop yields.
- Convolutional Neural Networks (CNNs) are increasingly vital for automated crop disease detection.
- Model training strategies and initial learning rates critically influence CNN performance.
Purpose of the Study:
- To evaluate the impact of different training strategies and initial learning rates on lightweight CNN models for wheat leaf disease identification.
- To identify the optimal combination for high accuracy and efficiency in disease detection.
Main Methods:
- Six training strategies (Adam, SGD, Adam + StepLR, SGD + StepLR, Warm-up + Cosine annealing + SGD, Warm-up + Cosine annealing + Adam) were tested.
- Three initial learning rates (0.05, 0.01, 0.001) were applied.
- Five lightweight CNN models (MobileNetV3, ShuffleNetV2, GhostNet, MnasNet, EfficientNetV2) were evaluated on wheat stripe rust, powdery mildew, and healthy wheat datasets.
Main Results:
- The MnasNet model, trained with SGD + StepLR and an initial learning rate of 0.001, achieved the highest recognition accuracy of 98.65%.
- This combination improved accuracy by 1.1% compared to fixed learning rate strategies.
- The MnasNet model demonstrated a compact parameter size of 19.09 M.
Conclusions:
- The MnasNet model, with the SGD + StepLR training strategy and 0.001 initial learning rate, is highly effective for automated wheat leaf disease identification.
- Its efficiency and accuracy make it suitable for deployment on mobile devices for real-time crop monitoring.
