近日，华中农业大学工学院柑橘全程机械化团队研究成果以“A generalization and lightweight recognition for citrus fruit harvesting based on improving YOLOv8”在Computers and Electronics in Agriculture发表。研究针对果实目标识别与定位需求，提出一款基于改进YOLOv8的轻量化柑橘果实通用识别模型“YOLOv8n-Light”，通过分析柑橘类果实在复杂果园背景中的成像规律，设计泛化和轻量化算法架构，实现橙、橘、柚三种柑橘大宗品种的高效识别，为柑橘采摘机器人的推广应用提供有力的技术支撑。
柑橘作为全球产量和种植面积居首的水果，其种植区多为山地丘陵，传统人工采摘面临劳动力短缺、成本高、效率低等问题，机械化、智能化采摘成为产业发展迫切需求。而精准的果实目标识别与定位，是柑橘采摘机器人实现高效作业的核心前提。
针对现有识别模型多面向单一柑橘品种、计算量高且难以适配嵌入式边缘设备的痛点，研究团队围绕模型轻量化与泛化性两大核心目标对YOLOv8进行优化：将骨干网络替换为轻量级ShuffleNetV2，大幅降低计算复杂度；构建基于共享卷积层的LightWeight Head检测头，减少参数冗余的同时提升跨品种特征提取能力；重新设计融合SE注意力机制的特征金字塔网络，强化复杂环境下的特征融合效果；整合MPDIoU与Focaler-IoU优势优化CIoU损失函数，实现检测精度与召回率的精准平衡，更适配采摘机器人的作业需求。模型性能验证结果显示，该模型能够同时识别橙、橘、柚三种果实，且与原始YOLOv8模型相比，计算量降低58%，工作站CPU推理速度提升45.9%（达15.35FPS），树莓派4B推理速度提升38.2%（达0.76FPS），在保证识别精度的同时，实现了模型轻量化与推理速度的双重提升。
研究团队将算法应用于多轴机械臂采摘作业控制，橙、橘、柚果实中心坐标的X、Y、Z方向平均误差均控制在10mm内，不同姿态下的半径误差均小于6mm，达到毫米级定位精度；果园实地开展的312次果实采摘测试充分验证了泛化模型与机械臂运动控制的高兼容性与实用性。
该研究提出的YOLOv8n-Light模型，既通过泛化算法设计突破了多品种目标的同时识别，又利用轻量化架构减轻了作业机具边缘设备的算力需求，为低成本、高效率柑橘果实采摘机具研发推广提供了技术支撑，也可作为多目标识别算法的研究参考。
华中农业大学工学院硕士研究生马京奥为论文第一作者，刘洁副教授为论文通讯作者。华中农业大学园艺林学学院伍小萌教授、解凯东副教授指导了试验设计和数据采集工作。该研究得到国家重点研发计划、国家柑橘产业技术体系和国家外国专家项目资助。
论文链接：https://doi.org/10.1016/j.compag.2026.111599
【英文摘要】
For efficient performance of selective harvesting robots, accurate target recognition and locating are the prerequisite and foundation. Therefore, the adaptability of fruit target recognition algorithms, both for the multiple fruit species and for the embedded edge computing platforms, has the potential to expand the application scope of picking robots. Since the citrus fruits have diverse morphology and a wide distribution in orchards, a lightweight citrus fruit recognition model based on improved YOLOv8 was designed to recognize the oranges （Citrus sinensis）， tangerines （Citrus reticulata） and pomelos （Citrus maxima） fruit on trees with less computation and tested on a picking robot arm by integrating with manipulator control programs. To improve the adaptability, the backbone network was replaced with lightweight ShuffleNetV2, a LightWeight Head detection head based on the concept of shared convolutional layers was constructed and the feature pyramid network （FPN） was redesigned into which SEAttention （Squeeze-Excitation Attention） was incorporated while integrating the advantages of MPDIoU and Focaler-IoU to construct an optimized CIoU loss function. With 7:1:2 as the ratio of training, validation and test sets, 2,250 images of orange, tangerine and pomelo fruit on the tree were involved in the training and test of the improved model. The weighted-average recognition precision, recall and mean average precision （mAP） of the test set were 92.5%, 82.0% and 89.4% for three species while the mAP values for independent
orange, tangerine and pomelo test sets were 92.6%, 86.9% and 88.7%, respectively. Compared with the original YOLOv8 model, the improved model reduced the computational load by 58%, with the inference speed reaching 15.35 FPS on a workstation CPU （an improvement of 45.9%） and 0.76 FPS on a Raspberry Pi 4B （an improvement of 38.2%）。 The localization experiment results showed that, in the manipulator base coordinate system, the average errors of the predicted center coordinates for oranges, tangerines and pomelos were within 10 mm in the X, Y and Z directions and the radius errors under different postures were all within 6 mm. To verify the practicality of the improved algorithm, 312 fruit grasping tests were conducted in orchards. The results showed that, for oranges, tangerines and pomelos, the recognition success rates were 93.5%, 93.6% and 88.3%, the localization success rates were 90.2%, 88.1% and 83.8% and the grasping success rates were 88.0%, 76.1% and 68.5%, respectively. The proposed YOLOv8n-Light model has provided a technical basis for citrus fruit picking operations and references for the design and improvement of fruit recognition algorithms for diverse fruit species.
日期：2026-10-02
