A lightweight intelligent grading method for lychee anthracnose based on improved YOLOv12

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Anthracnose significantly degrades lychee quality, necessitating efficient detection methods.
• Manual grading is inefficient and subjective, hindering effective quality management.
• Developing accurate yet lightweight AI models for disease detection is crucial for automated sorting.

Purpose of the Study:

• To develop a lightweight AI model for rapid, non-destructive detection and grading of lychee anthracnose.
• To improve upon existing detection frameworks by balancing high accuracy with reduced computational complexity.
• To provide a deployable solution for automated post-harvest lychee quality assessment.

Main Methods:

• Proposed LycheeGuard-Lite, a lightweight model based on the YOLOv12 framework.
• Incorporated C3k2_Light modules with depthwise separable convolutions and a C2PSA attention mechanism.
• Utilized a weighted convolution strategy (wConv2D) to enhance feature extraction and reduce complexity.
• Evaluated the model on a dataset of 14,576 images across two lychee varieties and three severity levels.

Main Results:

• Achieved 99.4% mAP50 detection accuracy for lychee anthracnose.
• Reduced model parameters by 12.8% (2.19M total).
• Decreased computational cost by 29.3% (4.1 GFLOPs).

Conclusions:

• LycheeGuard-Lite offers a highly accurate and computationally efficient solution for automated lychee disease recognition.
• The lightweight design makes the model suitable for practical deployment in post-harvest sorting systems.
• This research provides a valuable tool for improving lychee quality management and reducing food waste.