Abstract: Tillering number, a pivotal agronomic trait in wheat breeding programs, plays a crucial role in determining yield potential. However, traditional field-based tiller counting methods are plagued by inefficiencies, including labor-intensive manual assessments, subjective judgments, and challenges posed by dense tiller distribution, mutual occlusion, and diverse environmental conditions. These limitations not only hinder accurate phenotyping but also result in high operational costs and low reproducibility across trials. Moreover, human error and variability among observers further compromise data reliability. Consequently, the lack of precise, high-throughput tiller quantification impedes genetic progress and optimal management decisions. To address the limitations of manual phenotyping, there is an urgent need for automated, objective, and scalable solutions. Technologies such as image-based phenotyping coupled with machine learning can enhance measurement accuracy and reduce labor demands in breeding and agronomic research. Consequently, this study proposes YOLO-TN (YOLO for Tiller Number), which is a lightweight detection model designed to estimate wheat tiller counts from post-harvest drone imagery by identifying stem cross-sections. During field experiments, drones captured high-resolution images across diverse breeding materials and lighting conditions, which were subsequently standardized to a 640×640 pixel resolution. To optimize the model for dense, small-object detection within resource-constrained environments, the YOLO-TN architecture incorporates three primary enhancements. First, a shallow detection head, referred to as SmallHead, preserves fine-grained features from early-stage maps to mitigate information loss in deeper layers. Second, a global attention mechanism (GAM) is integrated to improve the model's sensitivity to tillering regions. This module dynamically evaluates and jointly models the importance of both channel-wise and spatial dimensions, thereby strengthening the perception of critical targets while suppressing interference from background clutter such as crop residues and weeds. Finally, to achieve a lightweight profile, RepViTBlocks are incorporated into the backbone network. These blocks employ reparameterized convolutions to significantly reduce computational overhead and parameter counts while simultaneously enhancing the joint representation of local details and global contextual information. Field experiments demonstrated the superior performance of the YOLO-TN model on the wheat tiller dataset. Field breeding experiments demonstrated that the YOLO-TN model exhibited exceptional performance on the tiller dataset of breeding wheat. At the individual plant level, YOLO-TN achieved an mAP@0.5 of 93.4% and an F1 score of 90.2%, marking improvements of 2.7 and 1.6 percentage points, respectively, over the YOLO11n. The coefficient of determination between the model’s predicted total tiller count and the manually annotated total tiller count was 0.92, with an root mean square error of 2.79. Additionally, at the plot scale, the coefficient of determination between the model’s predicted total tiller count and the total tiller count obtained from manual field surveys was 0.74, with an root mean square error of 4.59. Meanwhile, the model's parameter count and computational complexity are 3.48M and 11.3, respectively, representing reductions of 63% and 46.9% compared to YOLO11s. The optimized YOLO-TN model exhibits lightweight architecture, high precision, and high-throughput capabilities, enabling efficient tiller number acquisition in field breeding scenarios using drone technology. Its potential for large-scale wheat phenotyping applications underscores its significance in accelerating breeding programs. This study provides a reliable and scalable solution for tiller number assessment, contributing to the advancement of precision agriculture in wheat breeding.