Abstract
Anthracnose is one of the primary diseases leading to quality deterioration in lychee. Traditional manual grading methods suffer from low efficiency and high subjectivity. To achieve rapid, non-destructive detection and intelligent grading of lychee anthracnose, while addressing the challenge of balancing high accuracy and lightweight design in detection models, this study proposes a lightweight improved model named LycheeGuard-Lite based on the YOLOv12 framework. By introducing the C3k2_Light module reconstructed with depthwise separable convolutions, a dual-path C2PSA attention mechanism (position-channel dual-path attention), and the wConv2D weighted convolution strategy, the model enhances lesion feature extraction capability while reducing computational complexity.Evaluation was performed on a self-built dataset comprising 14, 576 images of two dominant lychee varieties ('Feizixiao' and 'Baitangying') collected under multiple lighting conditions and annotated with three severity levels (Mild, Moderate, Severe). The results demonstrate that the model maintains 99.4% mAP50 detection accuracy while reducing its number of parameters to 2.19M (a 12.8% decrease) and computational cost to 4.1 GFLOPs (a 29.3% reduction).This research provides a lightweight and deployable algorithmic foundation for automated lychee disease recognition and intelligent grading, offering practical engineering value for post-harvest fruit sorting and quality management.