Classification and Detection of Tea Based on Dynamic Sparse Attention

doi:10.13304/j.nykjdb.2024.0025

Abstract

Abstract:

In order to effectively address the challenge of precisely detecting and selecting tea leaves with a tea-picking robot， an improved YOLOv7 algorithm based on double-layer routing dynamic sparse attention mechanism and FasterNet was proposed to realize the classification and detection of fresh tea leaves. The algorithm implemented involves replacing the original network’s structure with PConv and FasterNet methodologies. This replacement was aimed at decreasing the number of floating-point operations and enhancing their efficiency. Additionally， a dynamic sparse attention mechanism， which was based on a double-layer routing approach， was incorporated into the neck layer. This addition ensured greater flexibility in computing allocation and content perception. To expedite convergence， enhance regression accuracy， and minimize false detection， the loss function was substituted with efficient intersection over union（EIoU）during the detection process. The results showed that， compared with the original YOLOv7， the model generated by the improved algorithm improved the accuracy by 4.8 percentage point， the recall rate by 5.3 percentage point， the balance score by 5.0 percentage point， and the mean average precision（mAP） value by 2.6 percentage point. In the external verification， the number of floating-point operations was reduced by 15.1 G， the frame per second was increased by 5.52%， and the mAP value was increased by 2.4 percentage point. The improved model could not only classify and detect fresh tea leaves efficiently and accurately， but also had the characteristics of high recognition rate， low computation and fast detection. Above results laid a foundation for the realization of tea picking robot in Yunnan plateau.

Key words: tea, vision transformer with Bi-level routing attention, accurate detection, FasterNet, YOLOv7

摘要：

为解决采茶机器人对茶叶的精准检测和采摘问题，提出一种基于双层路由动态稀疏注意力机制和FasterNet改进的YOLOv7算法，以实现对茶叶鲜叶的分类检测。该算法通过PConv和FasterNet替换原有网络结构，减少浮点运算的数量、提升浮点运算效率；在neck层加入基于双层路由的动态稀疏注意力机制，使计算分配和内容感知更灵活；将损失函数替换为EIoU（efficient intersection over union），加速收敛提高回归精度，减少检测过程中的误检。结果表明，改进算法生成的模型比YOLOv7在精确度上提升4.8个百分点，召回率提升5.3个百分点，平衡分数提高5.0个百分点，平均精度均值（mean average precision，mAP）提升2.6个百分点；且在外部验证中浮点运算数量降低15.1 G，每秒传输帧数提升5.52%，mAP提升2.4个百分点。改进后的模型不仅可以高效准确地对茶叶鲜叶进行分类检测，同时具备高识别率、低运算量和快速检测的特点。研究结果为云南高原山地采茶机器人的实现奠定了基础。

关键词: 茶叶, 双层路由动态稀疏注意力机制, 精准检测, FasterNet, YOLOv7

CLC Number:

TP391.41

Shihao ZHANG, Yuxin XIA, Wendou WU, Jin XIE, Xiao CHEN, Haotian SHI, Zongyu FAN, Baijuan WANG. Classification and Detection of Tea Based on Dynamic Sparse Attention[J]. Journal of Agricultural Science and Technology, 2025, 27(7): 111-121.

张世浩, 夏宇薪, 吴文斗, 谢瑾, 陈逍, 施皓天, 樊宗毓, 王白娟. 融合动态稀疏注意力的茶叶分类检测[J]. 中国农业科技导报, 2025, 27(7): 111-121.

Figures/Tables 10

Fig. 1 Sample image

Fig. 2 Improved YOLOv7 network architecture

Fig. 3 PConv structure

Fig. 4 FasterNet structure

Fig. 5 Bilevel routing attention （BRA） structure

Fig. 6 Loss function curve changeA：Improved YOLOv7 training set； B：YOLOv7 training set； C：Improved YOLOv7 validation set； D：YOLOv7 validation set

Fig. 7 Precision， recall and F1 score curveA：YOLOv7 precision； B：YOLOv7 recall； C：YOLOv7 F1 score； D：Improved YOLOv7 precision； E：Improved YOLOv7 recall； F：Improved YOLOv7 F1 score

Fig. 8 AP and mAP of different models

Fig. 9 Detection results of different models

Table 1 External validation parameters for comparing models

模型Model	平均精度均值mAP/%	浮点预算量FLOPs/G	每秒传输帧数FPS/Hz
改进型YOLOv7 Improved YOLOv7	95.1	89.7	46.08
YOLOv7	92.7	104.8	43.67
Faster-RCNN	81.3	377.6	7.03
SSD	89.6	268.4	18.97

References 31

[1]	刘玉飞,庞丹丹,蒋会兵,等.66份云南茶树种质生化成分的分析及特异种质筛选[J].南方农业学报,2021,52(3):693-699.
	LIU Y F, PANG D D, JIANG H B, et al.. Biochemical component analysis and specific resource selection of 66 accessions of tea germplasms in Yunnan [J]. J. South. Agric.,2021, 52(3): 693-699.
[2]	杨希,李玉川,罗倩倩,等.云南高山普洱茶渥堆过程滋味和汤色品质的动态变化研究[J].食品安全质量检测学报,2023,14(7):218-225.
	YANG X, LI Y C, LUO Q Q, et al.. Dynamic changes of taste and brewing color quality during the pile-fermentation process of Yunnan alpine Puerh tea [J]. J. Food Saf. Qual., 2023,14(7): 218-225.
[3]	冷彦,王静,吴文斗,等.高压脉冲电场预处理普洱熟茶对SD大鼠体内活性氧(ROS)水平的影响[J].西南农业学报,2021,34(6):1208-1214.
	LENG Y, WANG J, WU W D, et al.. Effect of pre-treatment of Pu’er tea by high voltage pulse electric field on reactive oxygen(ROS) levels in SD mice [J].Southwest China J. Agric. Sci., 2021,34(6):1208-1214.
[4]	朱智惟,单建华,余贤海,等.基于YOLOv5s的番茄采摘机器人目标检测技术[J].传感器与微系统,2023,42(6):129-132.
	ZHU Z W, SHAN J H, YU X H,et al..Target detection technology of tomato picking robot based on YOLOv5s [J].Transducer Microsyst. Technol., 2023,42(6):129-132.
[5]	刘思幸,李爽,缪宏,等.基于YOLOv3不同场景辣椒采摘机器人识别定位研究[J].农机化研究, 2024, 46(2): 38-43.
	LIU S X, LI S, MIAO H,et al..Research on recognition and location of pepper picking robot based on YOLOv3 in different scenarios [J]. J. Agric. Mech. Res., 2024, 46(2): 38-43.
[6]	金寿祥,周宏平,姜洪喆,等.采摘机器人视觉系统研究进展[J].江苏农业学报, 2023, 39(2): 582-595.
	JIN S X, ZHOU H P, JIANG H Z, et al.. Research progress on visual system of picking robot [J]. Jiangsu J. Agric. Sci., 2023,39(2): 582-595.
[7]	汪方斌,金蓄,朱达荣,等.基于RGB颜色空间的红外偏振人脸识别[J].激光与光电子学进展, 2022, 59(12): 535-543.
	WANG F B, JIN X, ZHU D R, et al.. Infrared polarized face recognition based on RGB color space [J]. Laser Optoelectron. Prog., 2022, 59(12): 535-543.
[8]	于明,李若曦,阎刚,等.基于颜色掩膜网络和自注意力机制的叶片病害识别方法[J].农业机械学报,2022,53(8):337-344.
	YU M, LI R X, YAN G, et al.. Crop diseases recognition method via fusion color mask and self-attention mechanism [J].Trans. Chin. Soc. Agric. Mach., 2022, 53(8):337-344.
[9]	肖冬娜,周忠发,尹林江,等.融合颜色指数与空间结构的喀斯特山地火龙果单株识别[J].激光与光电子学进展,2022,59(10):479-493.
	XIAO D N, ZHOU Z F, YIN L J, et al.. Identification of single plant of Karst Mountain pitaya by fusion of color index and spatial structure [J]. Laser Optoelectron. Prog., 2022, 59(10): 479-493.
[10]	王梦妮,顾寄南,王化佳,等.基于改进YOLOv5s模型的茶叶嫩芽识别方法[J].农业工程学报,2023,39(12):150-157.
	WANG M N, GU J N, WANG H J, et al.. Method for identifying tea buds based on improved YOLOv5s model [J].Trans. Chin. Soc. Agric. Eng., 2023, 39(12):150-157.
[11]	YAN L J, WU K H, LIN J, et al.. Identification and picking point positioning of tender tea shoots based on MR3P-TS model [J/OL]. Front. Plant Sci., 2022, 13:962391 [2023-12-26]. .
[12]	ZHANG S H, YANG H K, YANG C H, et al.. Edge device detection of tea leaves with one bud and two leaves based on ShuffleNetv2-YOLOv5-lite-E [J/OL]. Agronomy, 2023, 13(2): 577 [2023-12-26]. .
[13]	WANG C Y, BOCHKOVSKIY A, LIAO H Y M. YOLOv7:trainable bag-of-freebies sets new state-of-the-art for real-time object detectors [C]// 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023:7464-7475.
[14]	CHEN J R, KAO S H, HE H, et al.. Run, don’t walk:chasing higher FLOPS for faster neural networks [C]// 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023:12021-12031.
[15]	ZHU L, WANG X J, KE Z H, et al.. BiFormer:vision transformer with bi-level routing attention [C]// 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023:10323-10333.
[16]	杨世忠,王瑞彬,高升,等.基于YOLOv5的草莓轻量化网络检测模型[J].国外电子测量技术,2023,42(4):86-95.
	YANG S Z, WANG R B, GAO S,et al..YOLOv5-based lightweight network model for strawberry detection [J].Foreign Electron. Meas. Technol., 2023, 42(4): 86-95.
[17]	刘裕,赵保平,张述嘉,等.基于多尺度双路注意力胶囊网络在水稻害虫识别中的应用[J].西南农业学报,2022,35(7):1573-1581.
	LIU Y, ZHAO B P, ZHANG S J,et al..Application for recognition of rice pest based on multi-scale dual-path attention capsule network [J].Southwest China J.Agric.Sci.,2022,35(7):1573-1581.
[18]	朱旭,马淏,姬江涛,等.基于Faster R-CNN的蓝莓冠层果实检测识别分析[J].南方农业学报,2020,51(6):1493-1501.
	ZHU X, MA H, JI J T, et al.. Detecting and identifying blueberry canopy fruits based on Faster R-CNN [J]. J. South.Agric., 2020, 51(6):1493-1501.
[19]	国家质量监督检验检疫总局,中国国家标准化管理委员会. 地理标志产品普洱茶: [S].北京:中国标准出版社,2008.
[20]	杨断利,王永胜,陈辉,等.复杂环境下蛋鸡个体识别与自动计数系统研究[J].农业机械学报,2023,54(6):297-306.
	YANG D L, WANG Y S, CHEN H,et al..Individual identification and automatic counting system of laying hens under complex environment [J]. Trans. Chin. Soc. Agric. Mach.,2023, 54(6): 297-306.
[21]	WU D L, JIANG S, ZHAO E L, et al.. Detection of Camellia oleifera fruit in complex scenes by using YOLOv7 and data augmentation [J/OL]. Appl. Sci., 2022, 12(22): 11318 [2023-12-26]. .
[22]	CHEN Z, LIU C, FILARETOV V F, et al.. Multi-scale ship detection algorithm based on YOLOv7 for complex scene SAR images [J/OL]. Remote Sens., 2023, 15(8): 2071 [2023-12-26]. .
[23]	ZHANG C, HU Z H, XU L W, et al.. A YOLOv7 incorporating the Adan optimizer based corn pests identification method [J/OL].Front. Plant Sci.,2023,14:1174556 [2023-12-26]. .
[24]	卢俊哲,张铖怡,刘世鹏,等.面向复杂环境中带钢表面缺陷检测的轻量级DCN-YOLO[J].计算机工程与应用,2023,59(15):318-328.
	LU J Z, ZHANG C Y, LIU S P, et al.. Lightweight DCN-YOLO for strip surface defect detection in complex environments [J].Comput. Eng. Appl., 2023, 59(15): 318-328.
[25]	刘雄彪,杨贤昭,陈洋,等.基于CIoU改进边界框损失函数的目标检测方法[J].液晶与显示,2023,38(5):656-665.
	LIU X B, YANG X Z, CHEN Y, et al.. Object detection method based on CIoU improved bounding box loss function [J]. Chin. J. Liq. Cryst. Disp., 2023, 38(5): 656-665.
[26]	彭红星,徐慧明,高宗梅,等.基于改进YOLOF模型的田间农作物害虫检测方法[J].农业机械学报,2023,54(4):285-294, 303.
	PENG H X, XU H M, GAO Z M, et al.. Insect pest detection of field crops based on improved YOLOF model [J]. Trans. Chin.Soc. Agric. Mach., 2023, 54(4):285-294, 303.
[27]	张旭辉,闫建星,麻兵,等.基于改进YOLOv5s的护帮板异常检测方法研究[J].工程设计学报,2022,29(6):665-675.
	ZHANG X H, YAN J X, MA B, et al.. Research on abnormal detection method of side guard based on improved YOLOv5s [J]. Chin. J. Eng. Des., 2022, 29(6):665-675.
[28]	王琳毅,白静,李文静,等.YOLO系列目标检测算法研究进展[J].计算机工程与应用,2023,59(14):15-29.
	WANG L Y, BAI J, LI W J, et al.. Research progress of YOLO series target detection algorithms [J]. Comput. Eng. Appl., 2023, 59(14): 15-29.
[29]	高新阳,魏晟,温志庆,等.改进YOLOv5轻量级网络的柑橘检测方法[J].计算机工程与应用, 2023, 59(11): 212-221.
	GAO X Y, WEI S, WEN Z Q, et al.. Citrus detection method based on improved YOLOv5 lightweight network [J]. Comput. Eng. Appl., 2023, 59(11):212-221.
[30]	HUANG L, HUANG W Z, GONG H, et al.. PEFNet:position enhancement faster network for object detection in roadside perception system [J]. IEEE Access, 2881, 11: 73007-73023.
[31]	YANG Z J, FENG H L, RUAN Y P, et al.. Tea tree pest detection algorithm based on improved Yolov7-tiny [J/OL]. Agriculture, 2023,13(5):1031 [2023-12-26]. .