Abstract: To address the problems of low leaf recognition accuracy and blurred edge extraction that hinder high-precision phenotypic analysis, a data-driven real-time monitoring method is proposed to improve the automation and precision of plant phenotyping. To this end, based on the mainstream object detection model YOLOv5s, a novel object detection model (YOLO-STR) was constructed by integrating the Swin Transformer and RepVGG modules. The model integrates a global attention mechanism with a lightweight convolutional structure to enhance feature re-presentation in complex leaf scenes and achieve efficient and precise leaf area calculation. Wild-type Arabidopsis thaliana (WT) and ATP receptor mutants ( p1k1/p2k2 ) were used as experimental materials. Their full-lifecycle image data were collected under exogenous ATP treatment to construct datasets for model training and validation. On this basis, the model architecture was systematically evaluated through ablation experiments on different module combinations and comparative tests with multiple mainstream detection algorithms, comprehensively assessing its recognition and segmentation performance in complex leaf scenarios. The results showed that the YOLO-STR model achieved quantitative performance metrics exceeding 0.985 in automatic leaf segmentation, with an average accuracy (mAP@0.5) of 98.5%, significantly outperforming five representative models, including YOLOv7, YOLOv5s, and YOLOv3, with mAP improvements of 4.3%-23.6%. Ablation experiments further confirmed the synergistic role of the Swin Transformer and RepVGG modules in enhancing feature expression and model lightweighting: the Swin Transformer improved recognition accuracy by 7.3%, while the RepVGG reduced parameter count by 2.9 M. By integrating Gaussian area computation with camera geometric calibration, the method further enabled dynamic monitoring of Arabidopsis thaliana leaf growth. The monitoring results revealed that exogenous ATP treatment significantly inhibited the growth of wild-type plants, but had no obvious impact on mutant lines. In conclusion, the YOLO-STR model demonstrated high accuracy and strong real-time performance in automatic leaf recognition and dynamic area monitoring. It effectively addressed the challenge of insufficient recognition accuracy in dense leaf scenes and provides a reliable tool for quantitative analysis of plant growth rate and for investigating the physiological effects of exogenous ATP.