基于SOLOv2改进的苹果采摘机器人实例分割算法

doi:10.3969/j.issn.1004-1524.20250029

浙江农业学报 ›› 2026, Vol. 38 ›› Issue (3): 588-599.DOI: 10.3969/j.issn.1004-1524.20250029

基于SOLOv2改进的苹果采摘机器人实例分割算法

张郭¹(), 周庆辉²^,^*(), 何胜喜³

^1. 重庆科创职业学院智能制造与机器人学院, 重庆 402160
^2. 重庆大学机械传动国家重点实验室, 重庆 400044
^3. 重庆长安汽车股份有限公司, 重庆 400023

收稿日期:2025-01-06 出版日期:2026-03-25 发布日期:2026-04-17
作者简介:^*周庆辉,E-mail:zhouqinghui1985666@163.com
张郭,研究方向为自动化控制。E-mail:zhangguo1982@yeah.net
通讯作者: 周庆辉
基金资助:
国家自然科学基金(52005057)

An improved instance segmentation algorithm for apple picking robots based on SOLOv2

ZHANG Guo¹(), ZHOU Qinghui²^,^*(), HE Shengxi³

^1. College of Intelligent Manufacturing and Robotics, Chongqing College of Science and Creation, Chongqing 402160, China
^2. State Key Laboratory of Mechanical Transmission, Chongqing University, Chongqing 400044, China
^3. Chongqing Chang’an Automobile Co., Ltd., Chongqing 400023, China

Received:2025-01-06 Published:2026-03-25 Online:2026-04-17
Contact: ZHOU Qinghui

摘要/Abstract

摘要：

为解决现有苹果实例分割算法中存在的目标尺度多变、易受背景噪声干扰影响分割精度,以及推理速度过慢的问题,提出一种基于SOLOv2改进的苹果实例分割算法AE-SOLOv2。首先,采用StarBlock(星形卷积)优化特征提取网络,提升推理速度;然后,设计一种增强型双向特征金字塔网络(E-BiFPN),以额外的跨层连接促进高层语义信息与低层几何细节信息的充分融合,解决因分割目标尺度不一导致的精度下降问题;此外,引入注意力特征融合模块,通过综合多层次与多分支的特征信息,并结合自适应分组策略及全局与局部特征的协同处理,显著增强算法抗背景噪声干扰的能力。结果表明,改进后的算法能够有效提高苹果的实例分割精确率,在验证集上交并比为0.5时的平均精确率均值(mAP₅₀)达到91.2%,单张图像推理耗时12.3 ms,可满足苹果实例分割的实时性需求。此外,采摘实验表明,即便在计算资源受限的苹果采摘机器人场景下,AE-SOLOv2的各项性能仍表现出色,可为农业智能化发展提供一种新的解决思路。

关键词: SOLOv2网络, 实例分割, 采摘机器人, 注意力机制

Abstract:

To address the issues prevalent in existing apple instance segmentation algorithms, including variable object scales, susceptibility to background noise which degrades segmentation accuracy, and excessively slow inference speed, an improved SOLOv2-based apple instance segmentation algorithm, termed AE-SOLOv2, is proposed in this paper. First, the StarBlock structure is adopted to optimize the feature extraction network, thereby improving inference efficiency. Subsequently, an enhanced bidirectional feature pyramid network (E-BiFPN) is designed, where additional cross-layer connections facilitate adequate fusion of high-level semantic information and low-level geometric detail information, mitigating the accuracy degradation caused by inconsistent scales of segmentation targets. Furthermore, an attentive feature fusion module is introduced to integrate multi-level and multi-branch feature information. Combined with an adaptive grouping strategy and collaborative processing of global and local features, this module significantly enhances the anti-interference capability of the algorithm against background noise. Experimental results demonstrate that the improved algorithm effectively enhances the instance segmentation precision for apples. On the validation dataset, the mean average precision at an intersection-over-union threshold of 0.5 (mAP₅₀) reaches 91.2%, and the inference time per image is 12.3 ms, which satisfies the real-time requirements for apple instance segmentation. In addition, picking experiments verify that AE-SOLOv2 exhibits superior overall performance even in apple-picking robot scenarios with constrained computational resources, providing a novel solution for the development of agricultural intelligence.

Key words: SOLOv2 network, instance segmentation, picking robot, attention mechanism

中图分类号:

TP391.41
S24

张郭, 周庆辉, 何胜喜. 基于SOLOv2改进的苹果采摘机器人实例分割算法[J]. 浙江农业学报, 2026, 38(3): 588-599.

ZHANG Guo, ZHOU Qinghui, HE Shengxi. An improved instance segmentation algorithm for apple picking robots based on SOLOv2[J]. Acta Agriculturae Zhejiangensis, 2026, 38(3): 588-599.

图/表 12

图1 数据集示例

Fig.1 Examples of dataset

图2 AE-SOLOv2网络结构 AFFM,注意力特征融合模块。

Fig.2 AE-SOLOv2 network structure AFFM, Attention feature fusion module.

图3 不同颈部网络结构示意图

Fig.3 Structure of different neck networks

图4 注意力特征融合模块的网络结构

Fig.4 Adaptive feature fusion module network structure

图5 模型在训练集和验证集上的损失曲线

Fig.5 Loss curves of the model on the training set and validation set

表1 消融实验结果

Table 1 Results of ablation experiment

改进点 Improvements	mAP₅₀/%	Params/10⁶	FLOPs/10⁹	t/ms
—	87.3	46.6	179.5	24.7
StarBlock	87.4	21.6	81.9	11.9
StarBlock+E-BiFPN	89.0	21.7	82.4	11.9
StarBlock+	91.2	22.2	83.2	12.3
E-BiFPN+AFFM

图6 SOLOv2与AE-SOLOv2模型的P(精确率)-R(召回率)曲线

Fig.6 P(precision)-R(recall) curve of SOLOv2 and AE-SOLOv2 models

表2 对比实验的结果

Table 2 Comparative experimental results

模型Model	mAP_S/%	mAP_M/%	mAP_L/%	mAP_50:95/%	mAP₅₀/%	mAP₇₅/%	Params/10⁶	FLOPs/10⁹	t/ms
Hybrid Task Cascade	39.0	57.8	80.5	42.5	88.6	58.4	87.6	298.0	59.8
Mask RCNN	38.4	57.9	79.5	42.0	87.5	57.8	63.4	235.0	48.5
Cascade Mask RCNN	38.5	57.7	80.1	42.2	88.0	58.1	77.3	275.0	52.9
YOLACT	38.2	57.0	79.5	41.6	86.7	57.2	35.3	68.4	26.0
RTMDet	37.0	56.3	78.0	40.4	84.1	55.5	10.2	21.6	10.6
SOLOv2	38.3	57.9	79.4	41.9	87.3	57.6	46.6	179.5	24.7
AE-SOLOv2	39.9	59.0	81.5	43.8	91.2	60.2	22.2	83.2	12.3

图7 可视化对比的结果

Fig.7 Visual contrast result

图8 混淆矩阵预测结果对比

Fig.8 Comparison of prediction results of confusion matrix a,Hybrid Task Cascade;b,Mask RCNN;c,Cascade-Mask RCNN;d,YOLACT;e,RTMDet;f,SOLOv2;g,AE-SOLOv2。

图9 采摘机器人的采摘过程

Fig.9 Picking process of picking robot

表3 采摘机器人的感知性能对比

Table 3 Comparison of perceptual performance of picking robot

模型 Model	正确识别数 Correct identifications	漏检数 Missed detections	误检数 False detections	t/ms
SOLOv2	74	10	3	112.4
AE-SOLOv2	80	6	1	58.5

参考文献 25

[1]	冯青春, 赵春江, 李涛, 等. 苹果四臂采摘机器人系统设计与试验[J]. 农业工程学报, 2023, 39(13): 25-33.
	FENG Q C, ZHAO C J, LI T, et al. Design and test of a four-arm apple harvesting robot[J]. Transactions of the Chinese Society of Agricultural Engineering, 2023, 39(13): 25-33.
[2]	苑进. 选择性收获机器人技术研究进展与分析[J]. 农业机械学报, 2020, 51(9): 1-17.
	YUAN J. Research progress analysis of robotics selective harvesting technologies[J]. Transactions of the Chinese Society for Agricultural Machinery, 2020, 51(9): 1-17.
[3]	赵亚楠, 邓寒冰, 刘婷, 等. 基于弱监督学习的玉米苗期植株图像实例分割方法[J]. 农业工程学报, 2022, 38(19): 143-152.
	ZHAO Y N, DENG H B, LIU T, et al. Instance segmentation method of seedling maize plant images based on weak supervised learning[J]. Transactions of the Chinese Society of Agricultural Engineering, 2022, 38(19): 143-152.
[4]	LIU X X, CHEN Y P, WEI M Q, et al. Building instance extraction method based on improved hybrid task cascade[J]. IEEE Geoscience and Remote Sensing Letters, 2021, 19: 3002005.
[5]	JIA W K, WEI J M, ZHANG Q, et al. Accurate segmentation of green fruit based on optimized mask RCNN application in complex orchard[J]. Frontiers in Plant Science, 2022, 13: 955256.
[6]	高倩, 诸德宏, 封浩. 基于深度学习的番茄识别与实例分割[J]. 软件导刊, 2023, 22(2): 75-80.
	GAO Q, ZHU D H, FENG H. Tomato recognition and instance segmentation based on deep learning[J]. Software Guide, 2023, 22(2): 75-80.
[7]	JIA W K, TIAN Y Y, LUO R, et al. Detection and segmentation of overlapped fruits based on optimized mask R-CNN application in apple harvesting robot[J]. Computers and Electronics in Agriculture, 2020, 172: 105380.
[8]	ZENG J X, OUYANG H, LIU M, et al. Multi-scale YOLACT for instance segmentation[J]. Journal of King Saud University: Computer and Information Sciences, 2022, 34(10): 9419-9427.
[9]	RONG J C, HU L, ZHOU H, et al. A selective harvesting robot for cherry tomatoes: design, development, field evaluation analysis[J]. Journal of Field Robotics, 2024, 41(8): 2564-2582.
[10]	倪纪鹏, 朱立成, 董力中, 等. 基于SwinS-YOLACT的番茄采摘机器人实时实例分割算法研究[J]. 农业机械学报, 2024, 55(10): 18-30.
	NI J P, ZHU L C, DONG L Z, et al. Real-time instance segmentation algorithm for tomato picking robot based on SwinS-YOLACT[J]. Transactions of the Chinese Society for Agricultural Machinery, 2024, 55(10): 18-30.
[11]	朱德利, 余茂生, 梁明飞. 基于SwinT-YOLACT的玉米果穗实时实例分割[J]. 农业工程学报, 2023, 39(14): 164-172.
	ZHU D L, YU M S, LIANG M F. Real-time instance segmentation of maize ears using SwinT-YOLACT[J]. Transactions of the Chinese Society of Agricultural Engineering, 2023, 39(14): 164-172.
[12]	许楠, 苑迎春, 耿俊, 等. 基于改进YOLOv8s的果实与叶片器官分割方法[J]. 农业工程学报, 2024, 40(15): 119-126.
	XU N, YUAN Y C, GENG J, et al. Segmenting fruit and leaf organ using improved YOLOv8s[J]. Transactions of the Chinese Society of Agricultural Engineering, 2024, 40(15): 119-126.
[13]	张勤, 庞月生, 李彬. 基于实例分割的番茄串视觉定位与采摘姿态估算方法[J]. 农业机械学报, 2023, 54(10): 205-215.
	ZHANG Q, PANG Y S, LI B. Visual positioning and picking pose estimation of tomato clusters based on instance segmentation[J]. Transactions of the Chinese Society for Agricultural Machinery, 2023, 54(10): 205-215.
[14]	刘祥, 田敏, 梁金艳. 基于RCH-UNet的新疆密植棉花图像快速分割及产量预测[J]. 农业工程学报, 2024, 40(7): 230-239.
	LIU X, TIAN M, LIANG J Y. Prediction of cotton yield densely planted in Xinjiang of China using RCH-UNet model[J]. Transactions of the Chinese Society of Agricultural Engineering, 2024, 40(7): 230-239.
[15]	TANG S X, XIA Z L, GU J N, et al. High-precision apple recognition and localization method based on RGB-D and improved SOLOv2 instance segmentation[J]. Frontiers in Sustainable Food Systems, 2024, 8: 1403872.
[16]	LU Y J, WEI W, LI P X, et al. A deep learning method for building façade parsing utilizing improved SOLOv2 instance segmentation[J]. Energy and Buildings, 2023, 295: 113275.
[17]	WANG X L, ZHANG R F, KONG T, et al. Solov2:dynamic and fast instance segmentation[EB/OL]. [2025-01-06]. https://proceedings.neurips.cc/paper/2020/file/cd3afef-9b8b89558cd56638c3631868a-Paper.pdf.
[18]	SHEN Q, ZHANG L, ZHANG Y X, et al. Distracted driving behavior detection algorithm based on lightweight StarDL-YOLO[J]. Electronics, 2024, 13(16): 3216.
[19]	王洪波, 杨永政, 谢志成, 等. 基于Res-Inception的农作物病虫害识别技术[J]. 江苏农业科学, 2024, 52(20): 181-189.
	WANG H B, YANG Y Z, XIE Z C, et al. Crop diseases and pests identification technology based on Res-Inception[J]. Jiangsu Agricultural Sciences, 2024, 52(20): 181-189.
[20]	张磊, 李熙尉, 燕倩如, 等. 基于改进YOLOv5s的综采工作面人员检测算法[J]. 中国安全科学学报, 2023, 33(7): 82-89.
	ZHANG L, LI X W, YAN Q R, et al. Personnel detection algorithm in fully mechanized coal face based on improved YOLOv5s[J]. China Safety Science Journal, 2023, 33(7): 82-89.
[21]	林知心, 郑玉棒, 马天宇, 等. 基于轻量级全连接张量映射网络的高光谱图像分类方法[J]. 电子学报, 2024, 52(10): 3541-3551.
	LIN Z X, ZHENG Y B, MA T Y, et al. Lightweight fully-connected tensorial mapping network for hyperspectral image classification[J]. Acta Electronica Sinica, 2024, 52(10): 3541-3551.
[22]	DENG C F, WANG M M, LIU L, et al. Extended feature pyramid network for small object detection[J]. IEEE Transactions on Multimedia, 2021, 24: 1968-1979.
[23]	文雪梅, 王兴建, 宗炜佳, 等. 基于改进YOLOv5s模型的流体包裹体检测算法及其应用[J]. 科学技术与工程, 2024, 24(33): 14122-14128.
	WEN X M, WANG X J, ZONG W J, et al. Fluid inclusion detection algorithm based on improved YOLOv5s model and its application[J]. Science Technology and Engineering, 2024, 24(33): 14122-14128.
[24]	HE W F, LI C Z, NIE X F, et al. Recognition and detection of aero-engine blade damage based on Improved Cascade Mask R-CNN[J]. Applied Optics, 2021, 60(17): 5124-5133.
[25]	LIU S, ZOU H X, HUANG Y Z, et al. ERF-RTMDet: an improved small object detection method in remote sensing images[J]. Remote Sensing, 2023, 15(23): 5575.

基于SOLOv2改进的苹果采摘机器人实例分割算法

An improved instance segmentation algorithm for apple picking robots based on SOLOv2

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 25

相关文章 11

编辑推荐

Metrics

本文评价

[1]	汤志永, 何义川, 汤智辉, 张超, 刘湛, 刘晓瑜. 果园采摘机械的研究现状与发展趋势[J]. 浙江农业学报, 2026, 38(1): 184-196.
[2]	郑航, 冯昊栋, 薛向磊, 叶云翔, 于健麟, 俞国红. 基于实例分割的叶菜垄间导航线提取算法研究[J]. 浙江农业学报, 2025, 37(3): 701-711.
[3]	李萌民, 刘朔, 欧阳宇, 张鹏. 基于YOLOv8n的高效轻量化柑橘叶片病害检测模型[J]. 浙江农业学报, 2025, 37(10): 2198-2208.
[4]	朱铭敏, 张国平, 谭建军, 孙玲姣, 朱黎, 焦洁. 基于YOLOv5s的轻量级茶叶嫩芽终端检测模型[J]. 浙江农业学报, 2024, 36(6): 1413-1424.
[5]	李大华, 孔舒, 李栋, 于晓. 基于改进SSD模型的柑橘叶片病害轻量化检测模型[J]. 浙江农业学报, 2024, 36(3): 662-670.
[6]	李颀, 李煜哲. 弱光条件下猪舍清洗目标检测[J]. 浙江农业学报, 2023, 35(9): 2240-2249.
[7]	李荣鹏, 买买提·沙吾提, 盛艳芳, 何旭刚. 基于CA-MobileNet-V2的核桃病害识别与应用[J]. 浙江农业学报, 2023, 35(12): 2977-2987.
[8]	王英允, 龙燕, 杨智优, 黄铝文. 基于改进U-Net网络的苹果叶部病害语义分割方法[J]. 浙江农业学报, 2023, 35(11): 2731-2741.
[9]	周品志, 裴悦琨, 魏冉, 张永飞, 谷宇. 基于YOLOV4模型的果园樱桃实时检测研究[J]. 浙江农业学报, 2022, 34(11): 2522-2532.
[10]	张宁, 吴华瑞, 韩笑, 缪祎晟. 基于多尺度和注意力机制的番茄病害识别方法[J]. 浙江农业学报, 2021, 33(7): 1329-1338.
[11]	居洪玲;姬长英*. 一种多用途采摘机器人末端执行器的设计[J]. , 2010, 22(3): 0-373.