To enhance the recognition efficiency and accuracy of honeysuckle in intelligent harvesting, this study proposes a honeysuckle recognition method based on the improved YOLOv11m model. Addressing the challenges of small object size, color similarity with branches and leaves, and occlusion, this paper integrates a Deformable-LKA attention mechanism and a BiFPN bidirectional feature pyramid structure into the YOLOv11m framework. These enhancements improve the model’s perceptual and feature fusion capabilities for small targets in complex scenes. A dataset of 2195 honeysuckle images collected from various angles, lighting conditions, and flowering stages was manually annotated and augmented to 4700 images. Evaluation metrics include mAP@0.5, Precision, Recall, FPS, parameter count, and computational complexity. Results show that the improved YOLOv11m-DB model achieves an mAP@0.5 of 71.2%, an improvement of 4.6% points over the original model. Recall improved by 3.2%, and the frame rate increased by 9.2 FPS, demonstrating a strong balance between accuracy and efficiency. Compared to other lightweight models such as YOLOv9, YOLOv8m, and YOLOv11s, the proposed model exhibits superior detection performance and real-time capability while maintaining low parameter and computational costs. This study provides an effective solution for the target recognition module of honeysuckle harvesting robots and offers technical insights for identifying other medicinal plants.