Abstract: To address the challenge of inaccurate weed growth point localization in laser weeding—primarily caused by small target size, irregular morphology, and mutual occlusion in cornfield environments—this study proposes an integrated method that combines an improved deep learning detector with a novel skeleton and density peak analysis framework for apical meristem localization (SDPL).First, to enhance the detection of small, irregular, and partially occluded weeds, a lightweight model named YOLOv11n-LBD was developed. Based on YOLOv11n, the model incorporates three key enhancements: a Lightweight Adaptive Extraction (LAE) module for efficient and adaptive feature encoding, a Bidirectional Feature Pyramid Network (BiFPN) for improved multi-scale feature fusion with learnable weights, and a DyHeadDCNv4 detection head that integrates dynamic attention and deformable convolution for refined spatial and scale-aware feature representation. These modifications collectively strengthen the model's capability to extract discriminative features under complex field conditions. Evaluated on a custom dataset containing corn seedlings and five common weed species, the improved model achieved a mean Average Precision (mAP) of 94.8% and a recall of 93.2%. Compared to the baseline YOLOv11n, its parameter count and computational complexity were reduced by 0.4 million and 0.8 GFLOPs, respectively, while maintaining a real-time inference speed of 79 frames per second. This demonstrates a favorable balance between accuracy and efficiency suitable for mobile or embedded deployment in field robots.To overcome the significant positional deviation between the bounding box geometric center and the actual biological growth point of a weed, the SDPL localization pipeline was introduced. The process begins by cropping the region of interest (ROI) based on the YOLOv11n-LBD detection output. A high-quality binary mask is then generated using Gaussian filtering for noise suppression, followed by Otsu's adaptive thresholding for robust segmentation, effectively isolating the plant from the background. Morphological opening and closing operations are applied to suppress noise and enhance connectivity. Subsequently, a dedicated skeleton optimization framework is implemented. The initial single-pixel skeleton is extracted using the Zhang-Suen parallel thinning algorithm, preserving the topological structure. To address artifacts such as jaggedness and spurious branches, the raw skeleton is smoothed and refined through B-spline curve fitting, resulting in a continuous and natural skeletal representation. Candidate points for the growth point are defined as the junction points of the refined skeleton branches. Finally, the precise growth point is identified by analyzing the local pixel density distribution within a circular neighborhood around each candidate. The point exhibiting the highest local density peak, corresponding to the most concentrated area of plant tissue near the stem base, is selected as the final localization result. Experimental validation on field images showed that the proposed integrated method achieved a growth point localization accuracy of 81.2%. This is an improvement of 3 and 14 percentage points respectively compared to the key point detection and bounding box center positioning methods. The results demonstrate that the SDPL method, built upon robust detection and advanced morphological-densitometric analysis, performs reliably for weeds with irregular shapes, under partial occlusion, and even when overlapping with crop plants—scenarios where geometry-based methods often fail. This work provides a precise, practical, and computationally efficient visual perception solution for vision-guided laser weeding robots, contributing to targeted and energy-efficient automated weed control in sustainable agriculture.