TDS-YOLO: a lightweight detection model for fine-grained segmentation of tea leaf diseases

Abstract

Introduction: Timely identification and precise segmentation of tea leaf diseases are essential for intelligent agricultural management. However, balancing lightweight deployment and high-precision segmentation remains challenging under uneven illumination, background interference, and subtle early-stage lesion textures in natural environments. Methods: We propose TDS-YOLO, a lightweight segmentation model based on the YOLOv11 framework. The model introduces three innovations: (1) C3K2_EViM_CGLU for global dependency modeling, (2) EfficientHead for lightweight pixel-level representation, and (3) C2PSA_Mona to enhance multi-scale texture perception. Results: Experiments on a diverse dataset of 4,933 images show that TDS-YOLO achieves state-of-the-art performance with only 2.53M parameters. It reaches an mAP@0.5 of 90.1% for both detection and segmentation, outperforming YOLOv11-seg and other mainstream models while maintaining an inference speed of 96 FPS. Discussion: The proposed approach provides an efficient and robust solution for real-time monitoring of tea diseases, supporting precision tea plantation management and broader smart digital agriculture applications.