Identification of leaf diseases in field crops based on improved ShuffleNetV2

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Accurate crop disease identification is crucial for maintaining crop yields.
• Existing methods often suffer from large model sizes and low accuracy in complex field environments.

Purpose of the Study:

• To develop a lightweight and accurate crop leaf disease recognition model.
• To address limitations of existing models, including large parameter counts and poor performance in complex backgrounds.

Main Methods:

• Proposed a novel lightweight model, REM-ShuffleNetV2, based on an improved ShuffleNetV2 architecture.
• Incorporated residual structures for better feature learning and an Efficient Dual Channel Attention (EDCA) module for enhanced cross-channel interaction.
• Introduced multi-scale feature extraction modules to improve lesion detection at various scales.

Main Results:

• REM-ShuffleNetV2 achieved 96.72% accuracy and 96.62% F1 score on a field crop leaf disease dataset.
• The model has 4.40M parameters, 9.65% fewer than the original ShuffleNetV2.
• Outperformed established models like DenseNet121, EfficientNet, and MobileNetV3 in both accuracy and parameter efficiency.

Conclusions:

• REM-ShuffleNetV2 offers a lightweight and accurate solution for crop leaf disease identification in complex field settings.
• The model's small size facilitates deployment on mobile devices for intelligent crop disease diagnosis.
• Provides a valuable reference for developing efficient AI solutions in precision agriculture.