CTB-YOLO: A Lightweight Framework for Robust Cotton Terminal Bud Detection in Complex Field Environments

<p dir="ltr">Cotton terminal buds are critical for regulating plant architecture and determining yield formation, therefore, their automated detection is essential for intelligent cotton field cultivation. However, the ultra-small size of cotton buds, combined with complex illumination conditions and cluttered field backgrounds, poses significant challenges to robust real-time detection. To address these challenges, we present CTB-YOLO, a lightweight detection framework that integrates three complementary, purpose-built modules: the MSCA-FPN module (to prevent small-target feature dilution), the wConv2d module (to enhance illumination robustness), and the C2PSA-EDFFN module (to resolve contextual deficiency under severe occlusion). This integrated design ensures the framework systematically enhances fine-grained feature representation and multi-scale feature fusion in terms of both efficiency and robustness, concurrently minimizing model complexity and computational footprint. On our challenging self-constructed cotton terminal bud dataset, CTB-YOLO achieves a mean Average Precision (mAP50) of 85.3%. This result represents a significant 2.3% improvement over the YOLOv11 baseline while concurrently reducing model parameters by 28.3% and computational costs (GFLOPs) by 34.9%. Critically, our framework’s performance is highly competitive with benchmarked models, including the computationally intensive Transformer-based RT-DETR, while demanding only 3.2% of its computational resources. By delivering the highest detection accuracy with notable efficiency, CTB-YOLO establishes a favorable position on the accuracy-efficiency Pareto frontier, demonstrating its suitability for real-world deployment on resource-constrained agricultural edge devices. Furthermore, the proposed model demonstrates superior robustness under challenging field scenarios such as strong sunlight and severe target occlusion, and exhibits strong generalization capabilities validated on the Global Wheat Detection dataset. This framework provides a practical and efficient solution for automated plant management in precision agriculture.</p>