基于YOLOv8-Swin Transformer模型的翻覆肉鸭实时检测

doi:10.3969/j.issn.1004-1524.20240333

浙江农业学报 ›› 2025, Vol. 37 ›› Issue (7): 1556-1566.DOI: 10.3969/j.issn.1004-1524.20240333

基于YOLOv8-Swin Transformer模型的翻覆肉鸭实时检测

吕胤春¹^,²(), 段恩泽², 朱一星², 郑霞², 柏宗春²^,³^,^*()

1.江苏大学农业工程学院,江苏镇江 212000
2.江苏省农业科学院农业设施与装备研究所,江苏南京 210014
3.农业农村部长江中下游设施农业工程重点实验室,江苏南京 210014

收稿日期:2024-08-05 出版日期:2025-07-25 发布日期:2025-08-20
作者简介:吕胤春(1997—),男,江苏高邮人,硕士,研究方向为畜禽养殖技术与装备。E-mail:18851730800@163.com
通讯作者: ^*柏宗春,E-mail:vipmaple@126.com
基金资助:
江苏省农业科技自主创新资金“江苏现代农业重大核心技术创新”类项目(CX〔22〕1008)

Real-time detection of overturned meat ducks based on YOLOv8-Swin Transformer model

LYU Yinchun¹^,²(), DUAN Enze², ZHU Yixing², ZHENG Xia², BAI Zongchun²^,³^,^*()

1. College of Agricultural Engineering, Jiangsu University, Zhenjiang 212000, Jiangsu, China
2. Institute of Agricultural Facilities and Equipment, Jiangsu Academy of Agricultural Sciences, Nanjing 210014, China
3. Key Laboratory of Protected Agriculture Engineering in the Middle and Lower Reaches of Yangtze River, Ministry of Agriculture and Rural Affairs, Nanjing 210014, China

Received:2024-08-05 Online:2025-07-25 Published:2025-08-20

摘要/Abstract

摘要：

针对规模化养殖场笼内肉鸭个体小、易被遮挡,且肉鸭翻覆目标检测方法不易在嵌入式端部署等问题,提出一种适用于Jetson Orin端部署的肉鸭翻覆行为检测方法,在准确检测翻覆肉鸭目标的同时,轻量化部署模型,提高检测效率。使用1 000幅翻覆肉鸭图像建立数据集,按8∶1∶1划分为训练集、测试集和验证集。利用深度学习网络提取肉鸭翻覆行为特征,构建肉鸭翻覆行为目标检测模型。使用Swin Transformer-tiny模块替换YOLOv8的主干网络,有效提升复杂环境下的小目标检测能力,通过对模型进行剪枝与量化以减轻模型的复杂度,同时保持精度,较好地平衡了模型的准确性和速度。将优化后的模型部署在嵌入式端,当置信度阈值设定为60时,YOLOv8n-Swin Transformer和YOLOv8s-Swin Transformer模型对肉鸭翻覆的识别平均准确率分别为96.0%和97.1%,识别误检率分别为2.7%和2.0%,单帧图像处理时间分别为6.8 ms和7.4 ms。

关键词: 机器视觉, 翻覆肉鸭识别, 笼养肉鸭, 深度学习, 设施养殖

Abstract:

To address the challenges of detecting small and easily occluded meat ducks in cages within large-scale farms, as well as the difficulty of deploying existing detection methods for overturned ducks on embedded devices, this study proposes a detection method for identifying overturned meat ducks suitable for deployment on Jetson Orin. This approach ensures accurate detection of overturned ducks while achieving lightweight model deployment and improved detection efficiency. A dataset of 1 000 images of overturned meat ducks was constructed and divided into training, testing, and validation sets in an 8∶1∶1 ratio. A deep learning network was employed to extract behavioral features of overturned ducks and build a target detection model. The Swin Transformer-tiny module was integrated to replace the backbone network of YOLOv8, significantly enhancing the detection capability for small targets in complex environments. Model pruning and quantization were applied to reduce computational complexity while maintaining accuracy, achieving a better balance between model precision and speed. When deploying the optimized models on embedded devices with a confidence threshold set to 60, the YOLOv8n-Swin Transformer and YOLOv8s-Swin Transformer models demonstrated average recognition accuracies of 96.0% and 97.1%, respectively, for detecting overturned meat ducks. Their false recognition rates were 2.7% and 2.0%, while the single-frame image processing times measured 6.8 ms and 7.4 ms, respectively.

Key words: machine vision, overturned meat duck recognition, caged meat duck, deep learning, facility breeding

中图分类号:

吕胤春, 段恩泽, 朱一星, 郑霞, 柏宗春. 基于YOLOv8-Swin Transformer模型的翻覆肉鸭实时检测[J]. 浙江农业学报, 2025, 37(7): 1556-1566.

LYU Yinchun, DUAN Enze, ZHU Yixing, ZHENG Xia, BAI Zongchun. Real-time detection of overturned meat ducks based on YOLOv8-Swin Transformer model[J]. Acta Agriculturae Zhejiangensis, 2025, 37(7): 1556-1566.

图/表 9

图1 巡检装备尺寸和摄像头与肉鸭笼具之间的相对位置

Fig.1 Inspection equipment size and relative position between camera and duck cage

图2 图像增强处理过程

Fig.2 Image enhancement process

图3 YOLOv8网络的架构 C,卷积层;U,上采样;Bbox,边界框;Cls,类别分数;C2F,增强特征提取;Conv,卷积层。下同。

Fig.3 Architecture of YOLOv8 network C, Convolutional layer; U, Upsample; Bbox, Bounding box; Cls, Classification score; C2F, Cross stage partial bottleneck with two convolutions; Conv, Convolutional layer. The same as below.

图4 Swin Transformer block的整体架构 LN,层标准化;W-MSA,基于局部窗口的多头自注意力机制;SW-MSA,移位窗口注意力机制;MLP,多层感知机。

Fig.4 Overall architecture of Swin Transformer block LN, Layer normalization; W-MSA, Window-based multi-head self-attention; SW-MSA, Shifted window multi-head self-attention; MLP, Muti-layer perceptron.

图5 窗口移位作用流程 a,输入图像;b,基于W-MSA的窗口分割;c,移动窗口;d,基于SW-MSA的窗口分割。

Fig.5 Flow of window moving action a, Input image; b, Window segmentation via W-MSA; c, Moving window effect; d, Window segmentation via SW-MSA.

图6 基于Swin Transformer改进的YOLOv8网络结构

Fig.6 Improved YOLOv8 network structure based on Swin Transformer

图7 边框损失值和分类损失值曲线

Fig.7 Curves of bounding box loss and classification loss

图8 精确率、召回率和平均精度均值曲线

Fig.8 Curves of precision, recall and mean average precision

表1 不同模型在不同置信度阈值下的肉鸭翻覆识别性能

Table 1 Performance of different models on meat duck turnover recognition under different confidence thresholds

模型 Model	置信度阈值 Confidence threshold	误检数 False detection	漏检数 Missed detection	识别平均准确率 Average recognition accuracy/%	识别误检率 False recognition rate/%	单帧图像处理时间 Single frame image processing time/ms
YOLOv5n	60	6	1	93.4	4.9	4.8
	90	3	9	88.3	8.7	4.8
YOLOv8n	60	5	1	94.3	4.2	4.4
	90	3	8	89.3	7.9	4.4
YOLOv8n-Swin Transformer	60	4	0	96.0	2.7	6.8
	90	1	7	92.1	5.6	6.8
YOLOv8s-Swin Transformer	60	3	0	97.1	2.0	7.4
	90	1	6	93.1	4.9	7.4

参考文献 23

[1]	黄智. 肉鸭规模化生态养殖关键技术[J]. 畜牧兽医科技信息, 2025(1): 211-213.
	HUANG Z. Key technologies of large-scale ecological breeding of meat ducks[J]. Chinese Journal of Animal Husbandry and Veterinary Medicine, 2025(1): 211-213. (in Chinese)
[2]	徐佳. 鸭传染性浆膜炎的临床症状及防治措施[J]. 家禽科学, 2024(2): 78-80.
	XU J. Clinical symptoms and preventive measures of duck infectious serositis[J]. China Poultry Science, 2024(2): 78-80. (in Chinese)
[3]	刘芝美. 肉鸭养殖中存在的问题及对策[J]. 养禽与禽病防治, 2013(9): 32-33.
	LIU Z M. Problems and countermeasures in meat duck breeding[J]. Poultry Husbandry and Disease Control, 2013(9): 32-33. (in Chinese)
[4]	刘又夫, 肖德琴, 周家鑫, 等. 水禽智能化养殖研究现状及发展趋势[J]. 智慧农业(中英文), 2023, 5(1): 99-110.
	LIU Y F, XIAO D Q, ZHOU J X, et al. Status quo of waterfowl intelligent farming research review and development trend analysis[J]. Smart Agriculture, 2023, 5(1): 99-110. (in Chinese with English abstract)
[5]	付友, 王成森, 郭瑞萍, 等. 肉鸭养殖环控智能化管理技术探讨[J]. 家禽科学, 2023(8): 59-61.
	FU Y, WANG C S, GUO R P, et al. Discussion on intelligent management technology of environmental control for meat duck breeding[J]. China Poultry Science, 2023(8): 59-61. (in Chinese)
[6]	唐瑜嵘, 沈明霞, 薛鸿翔, 等. 人工智能技术在畜禽养殖业的发展现状与展望[J]. 智能化农业装备学报(中英文), 2023(1): 1-16.
	TANG Y R, SHEN M X, XUE H X, et al. Development status and prospect of artificial intelligence technology in livestock and poultry breeding[J]. Journal of Intelligent Agricultural Mechanization, 2023(1): 1-16. (in Chinese with English abstract)
[7]	赵春江, 梁雪文, 于合龙, 等. 基于改进YOLO v7的笼养鸡/蛋自动识别与计数方法[J]. 农业机械学报, 2023, 54(7): 300-312.
	ZHAO C J, LIANG X W, YU H L, et al. Automatic identification and counting method of caged hens and eggs based on improved YOLO v7[J]. Transactions of the Chinese Society for Agricultural Machinery, 2023, 54(7): 300-312. (in Chinese with English abstract)
[8]	刘啸虎, 肖德琴, 刘又夫, 等. 基于Faster R-CNN和时序统计的肉鸭行为节律分析[J]. 中国家禽, 2023, 45(11): 95-104.
	LIU X H, XIAO D Q, LIU Y F, et al. Analysis on rhythmic behavior of meat ducks based on faster R-CNN and time-series statistics[J]. China Poultry, 2023, 45(11): 95-104. (in Chinese with English abstract)
[9]	马肄恒. 面向笼养肉鸭行为与死亡识别的自主巡检装备创制[D]. 杭州: 浙江科技大学, 2024.
	MA Y H. Development of autonomous inspection equipment for behavior and mortality detection in captive broiler ducks[D]. Hangzhou, 2024. (in Chinese with English abstract)
[10]	姜来, 王文娣, 霍晓静, 等. 死鸡识别机器人系统设计与试验[J]. 中国农机化学报, 2023, 44(8): 81-87.
	JIANG L, WANG W D, HUO X J, et al. Design and experiment of dead chicken recognition robot system[J]. Journal of Chinese Agricultural Mechanization, 2023, 44(8): 81-87. (in Chinese with English abstract)
[11]	LIU H W, CHEN C H, TSAI Y C, et al. Identifying images of dead chickens with a chicken removal system integrated with a deep learning algorithm[J]. Sensors, 2021, 21(11): 3579.
[12]	贾雁琳, 薛皓, 周子轩, 等. 基于红外热成像技术的笼内死鸡自动识别方法[J]. 河北农业大学学报, 2023, 46(3): 105-112.
	JIA Y L, XUE H, ZHOU Z X, et al. Automatic identification method for dead chicken in cage based on infrared thermal imaging technology[J]. Journal of Hebei Agricultural University, 2023, 46(3): 105-112. (in Chinese with English abstract)
[13]	YANG C C, CHAO K, CHEN Y R, et al. Simple multispectral image analysis for systemically diseased chicken identification[J]. Transactions of the ASABE, 2006, 49(1): 245-257.
[14]	JIANG P Y, ERGU D J, LIU F Y, et al. A review of yolo algorithm developments[J]. Procedia Computer Science, 2022, 199: 1066-1073.
[15]	HUSSAIN M. YOLO-v1 to YOLO-v8, the rise of YOLO and its complementary nature toward digital manufacturing and industrial defect detection[J]. Machines, 2023, 11(7): 677.
[16]	WANG H, ZHANG F, WANG L. Fruit classification model based on improved Darknet 53 convolutional neural network[C]// 2020 International Conference on Intelligent Transportation, Big Data & Smart City (ICITBS). January 11-12, 2020, Vientiane, Laos. IEEE, 2020: 881-884.
[17]	WATANABE N. Two types of graphite fluorides, (CF)n and (C2F)n, and discharge characteristics and mechanisms of electrodes of (CF)n and (C2F)n in lithium batteries[J]. Solid State Ionics, 1980, 1(1/2): 87-110.
[18]	MAHAREK A, ABOZEID A, ORBAN R, et al. SwinVid: enhancing video object detection using swin transformer[J]. Computer Systems Science and Engineering, 2024, 48(2): 305-320.
[19]	LI J P, YAN Y C, LIAO S C, et al. Local-to-global self-attention in vision transformers[EB/OL]. (2021-07-10)[2024-08-04]. https://arxiv.org/abs/2107.04735v1.
[20]	LIU Z, LIN Y T, CAO Y, et al. Swin transformer: hierarchical vision transformer using shifted windows[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). October 10-17, 2021, Montreal, QC, Canada. IEEE, 2021: 9992-10002.
[21]	ZHENG Z H, WANG P, LIU W, et al. Distance-IoU loss: faster and better learning for bounding box regression[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2020, 34(7): 12993-13000.
[22]	GEVORGYAN Z. SIoU loss: more powerful learning for bounding box regression[EB/OL]. (2022-05-25)[2024-08-04]. https://arxiv.org/abs/2205.12740v1.
[23]	ZHANG Z L, SABUNCU M R. Generalized cross entropy loss for training deep neural networks with noisy labels[J]. Advances in Neural Information Processing Systems, 2018, 32: 8792-8802.

基于YOLOv8-Swin Transformer模型的翻覆肉鸭实时检测

Real-time detection of overturned meat ducks based on YOLOv8-Swin Transformer model

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 9

参考文献 23

相关文章 15

编辑推荐

Metrics

本文评价

[1]	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.
[2]	郭秀明, 王大伟, 刘升平, 诸叶平, 刘晓辉, 林克剑, 王佳宇, 李非. 基于深度学习的近地面草原鼠洞识别计数关键问题研究与应用[J]. 浙江农业学报, 2024, 36(9): 2146-2154.
[3]	程嘉瑜, 陈妙金, 李彤, 孙奇男, 张小斌, 赵懿滢, 朱怡航, 顾清. 基于改进Faster-RCNN网络的无人机遥感影像桃树检测[J]. 浙江农业学报, 2024, 36(8): 1909-1919.
[4]	宁文楷, 李静, 沈晓东, 吴鑫, 李臻锋. 南瓜干燥过程中β-胡萝卜素的多源融合预测[J]. 浙江农业学报, 2023, 35(8): 1876-1887.
[5]	潘攀, 张建华, 郑晓明, 周国民, 胡林, 冯全, 柴秀娟. 深度学习在作物及其近缘种抗病性智能鉴定上的研究进展[J]. 浙江农业学报, 2023, 35(8): 1993-2012.
[6]	白卫卫, 赵雪妮, 罗斌, 赵薇, 黄硕, 张晗. 基于YOLOv5的小麦种子发芽检测方法研究[J]. 浙江农业学报, 2023, 35(2): 445-454.
[7]	朱世松, 马婉丽, 赵理山, 郑艳梅, 郑先波, 芦碧波. 基于改进的LinkNet的苹果叶片图像分割算法[J]. 浙江农业学报, 2023, 35(1): 202-214.
[8]	陈道怀, 汪杭军. 基于改进YOLOv4的林业害虫检测[J]. 浙江农业学报, 2022, 34(6): 1306-1315.
[9]	闫宁, 张晗, 董宏图, 康凯, 罗斌. 基于透射光和反射光图像同位分割的小麦品种识别方法研究[J]. 浙江农业学报, 2022, 34(3): 590-598.
[10]	李超, 李锋, 黄炜嘉. 基于并联卷积神经网络的水果品种识别[J]. 浙江农业学报, 2022, 34(11): 2533-2541.
[11]	张晴晴, 刘连忠, 宁井铭, 吴国栋, 江朝晖, 李孟杰, 李栋梁. 基于YOLOV3优化模型的复杂场景下茶树嫩芽识别[J]. 浙江农业学报, 2021, 33(9): 1740-1747.
[12]	鲍烈, 王曼韬, 刘江川, 文波, 明月. 基于卷积神经网络的小麦产量预估方法[J]. 浙江农业学报, 2020, 32(12): 2244-2252.
[13]	包晓敏, 盛家文. 粘虫板害虫自动识别计数研究[J]. 浙江农业学报, 2019, 31(9): 1516-1522.
[14]	王彦翔, 张艳, 杨成娅, 孟庆龙, 尚静. 基于深度学习的农作物病害图像识别技术进展[J]. 浙江农业学报, 2019, 31(4): 669-676.
[15]	杨国亮, 许楠, 康乐乐, 龚曼, 洪志阳. 基于参数指数非线性残差神经网络的脐橙病变叶片识别[J]. 浙江农业学报, 2018, 30(6): 1073-1081.