YOLO-lychee-advanced: an optimized detection model for lychee pest damage based on YOLOv11

We introduce YOLO-Lychee-advanced, a lightweight and high-precision detector for lychee stem-borer damage on fruit surfaces. Built on YOLOv11, the model incorporates (i) a C2f module with dual-branch residual connections to capture fine-grained features of pest holes ≤2 mm, (ii) a CBAM channel-spatial attention block to suppress complex peel-texture interference, and (iii) CIoU loss to tighten bounding-box regression. To mitigate illumination variance, we augment the original 3,061-image dataset to 9,183 samples by simulating direct/back-lighting and adopt a "pest-hole only" annotation strategy, which improves mAP50-95 by 18% over baseline. Experiments conducted on an RTX 3060 with a batch size of 32 and an input size of 416 × 416 pixels show YOLO-Lychee-advanced achieves 92.2% precision, 85.4% recall, 91.7% mAP50, and 61.6% mAP50-95, surpassing YOLOv9t and YOLOv10n by 3.4% and 1.7%, respectively, while maintaining 37 FPS real-time speed. Compared with the recent YOLOv9t and YOLOv10n baselines on the same lychee test set, YOLO-Lychee-advanced raises mAP50-95 by 3.4% and 1.7%, respectively. Post-processing optimization further boosts precision to 95.5%. A publicly available dataset and PyQt5 visualization tool are provided at https://github.com/Suxueping/Lychee-Pest-Damage-images.git.