Grape Leaf Cultivar Identification in Complex Backgrounds with an Improved MobileNetV3-Small Model

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Science

Background:

• Accurate grape leaf identification is crucial for viticulture management, breeding, and precision agriculture.
• Challenges include subtle morphological variations, background noise, and the need for computational efficiency in mobile applications.
• Existing methods struggle to balance accuracy and efficiency for practical field deployment.

Purpose of the Study:

• To develop an improved lightweight convolutional neural network (CNN) for accurate grape leaf variety recognition.
• To enhance feature extraction, multi-scale fusion, and feature distribution optimization for improved classification robustness.
• To provide a computationally efficient solution for mobile-based precision viticulture.

Main Methods:

• Proposed an improved lightweight CNN, ICS-MobileNetV3-Small (ICS-MS).
• Integrated a coordinate attention mechanism for enhanced spatial feature capture and noise suppression.
• Incorporated a multi-branch ICS-Inception structure for multi-scale feature fusion.
• Utilized a joint loss function to optimize feature space distribution and classification robustness.

Main Results:

• Achieved 96.53% recognition accuracy with only 1.17 million parameters on an eleven-variety grape leaf dataset.
• The ICS-MS model demonstrated significant improvements over the baseline MobileNetV3-Small.
• Precision, recall, and F1-scores were consistently near 96%, indicating a favorable efficiency-accuracy trade-off.

Conclusions:

• ICS-MS offers a practical and reliable approach for grape leaf identification, addressing key challenges in feature extraction and computational efficiency.
• The model's lightweight design and high accuracy make it suitable for mobile deployment in precision viticulture.
• This work provides a valuable tool to support intelligent management and decision-making in viticulture.