Accurate detection of cotton top buds is a key prerequisite for precision topping. However, because of the small size of cotton top buds and their high visual similarity to complex canopy backgrounds, high-efficiency detection in natural fields still faces challenges. To address these challenges, we propose an advanced lightweight, real-time, and accurate detection model, MC-RT-DETR. Specifically, we replace the original ResNet18 backbone network of RT-DETR-r18 with MobileNetV3, which significantly reduces the model parameters and computational complexity. Then, we integrate the C2f module into the feature fusion network, which effectively enhances the multiscale feature representation. We employ Wise-IoU to optimize bounding box regression, achieving high-IoU localization accuracy for small targets. Finally, strategies such as increasing the bounding box regression loss function and using MixUp and Copy-Paste data enhancement for decoupled optimization are combined, which improves the robustness of the model in the context of complex crowns. Experiments show that MC-RT-DETR achieves mAP@50 of 96.00% and mAP@50–95 of 83.59% while reducing the number of parameters and GFLOPs by 47.2% and 57.3%, respectively. Edge deployment on Jetson Orin Nano demonstrates that the model achieves a frame rate of 65.9 FPS, fully satisfying real-time field requirements.