Accurate estimates of cotton seedling abundance and spatial distribution support emergence assessment and field management. However, seedlings occupy few pixels in unmanned aerial vehicle (UAV) imagery, occur at high densities, and frequently overlap. This study proposes a High-Resolution Center-Guided Feature Fusion model (HCF-YOLO) that combines a P2 detection layer, Lightweight Multi-Scale Feature Fusion (LMSFF), and training-only Gaussian Center Guidance (GCG). The dataset contained 1040 images and 101,824 annotated cotton seedlings and was divided into 832 training, 104 validation, and 104 test images. On the main test split, HCF-YOLO11n achieved an mAP50 of 0.9354 and an mAP50-95 of 0.4431, improvements of 5.10 and 3.74 percentage points over YOLO11n, while reducing the parameter count and FLOPs by 14.18% and 14.63%. The performance ranking was consistent across three random seeds and was retained in the external-source evaluation. Under the pruning protocol, the 5%-pruned HCF-YOLO11n achieved an mAP50-95 of 0.4578 on the same 104-image test split while reducing parameters and FLOPs by 11.19% and 6.87%, respectively. TensorRT FP16 reduced network inference latency to 1.453 ms, and Raspberry Pi 5 deployment reached 6.67 FPS. Overall, HCF-YOLO balances accuracy, model size, and edge efficiency for UAV-assisted cotton seedling recognition and image-plane localization, although onboard validation remains necessary.