基于改进YOLACT的油茶叶片炭疽病感染严重程度分级模型

doi:10.12133/j.smartag.SA202402002

摘要/Abstract

摘要：

[目的/意义] 炭疽病（anthracnose）作为油茶生长过程中重要的病害，其严重程度的精准判定对于精准施药和科学管理具有重大意义。本研究提出了一种改进YOLACT（You Only Look At CoefficienTs）分级模型Camellia-YOLACT，旨在实现对油茶叶片炭疽病感染严重程度的自动、高效判定。 [方法] 首先在YOLACT主干网络部分使用Swin-Transformer来进行特征提取。Transformer架构的自注意力机制拥有全局感受野及移位窗口等特性，有效地增强了模型的特征提取能力；引入加权双向特征金字塔网络，融合不同尺度的特征信息，加强模型对不同尺度目标的检测能力，提高模型的检测精度；在激活函数的选择上，采用非线性能力更强的HardSwish激活函数替换原模型的ReLu激活函数。由于HardSwish在负值区域不是完全截断，对于输入数据中的噪声具有更高的鲁棒性，自然环境下的图像有着复杂的背景和前景信息，HardSwish的鲁棒性有助于模型更好地处理这些情况，进一步提升精度。 [结果和讨论] 采用迁移学习方式在油茶炭疽病感染严重程度分级数据集上进行实验验证。消融实验结果表明，本研究提出的Camellia-YOLACT模型的mAP₇₅为86.8%，较改进前提升5.7%；mAP_all为78.3%，较改进前提升2.5%；mAR为91.6%，较改进前提升7.9%。对比实验结果表明，Camellia-YOLACT在精度和速度方面表现均好于SOLO（Segmenting Objects by Locations），与Mask R-CNN算法相比，其检测速度提升了2倍。在室外的36组分级实验中进一步验证了Camellia-YOLACT模型的性能，其对油茶炭疽病严重程度的分级正确率达到了94.4%，K值平均绝对误差为1.09%。 [结论] 本研究提出的Camellia-YOLACT模型在油茶叶片和炭疽病病斑分割上具有较高的精度，能够实现对油茶炭疽病严重程度的自动分级，为油茶病害的精准防治提供技术支持，进一步推动油茶炭疽病诊断的自动化和智能化。

关键词: 油茶, 叶部病害, 炭疽病, BiFPN, YOLACT, Transformer, 深度学习

Abstract:

[Objective] Camellia oleifera is one of the four major woody oil plants in the world. Diseases is a significant factor leading to the decline in quality of Camellia oleifera and the financial loss of farmers. Among these diseases, anthracnose is a common and severe disease in Camellia oleifera forests, directly impacting yields and production rates. Accurate disease assessment can improve the prevention and control efficiency and safeguarding the farmers' profit. In this study, an improved You Only Look at CoefficienTs (YOLACT) based method was proposed to realize automatic and efficient grading of the severity of Camellia oleifera leaf anthracnose. [Methods] High-resolution images of Camellia oleifera anthracnose leaves were collected using a smartphone at the National Camellia oleifera Seed Base of Jiangxi Academy of Forestry, and finally 975 valid images were retained after a rigorous screening process. Five data enhancement means were applied, and a data set of 5 850 images was constructed finally, which was divided into training, validation, and test sets in a ratio of 7:2:1. For model selection, the Camellia-YOLACT model was proposed based on the YOLACT instance segmentation model, and by introducing improvements such as Swin-Transformer, weighted bi-directional feature pyramid network, and HardSwish activation function. The Swin Transformer was utilized for feature extraction in the backbone network part of YOLACT, leveraging the global receptive field and shift window properties of the self-attention mechanism in the Transformer architecture to enhance feature extraction capabilities. Additionally, a weighted bidirectional feature pyramid network was introduced to fuse feature information from different scales to improve the detection ability of the model for objects at different scales, thereby improving the detection accuracy. Furthermore, to increase the the model's robustness against the noise in the input data, the HardSwish activation function with stronger nonlinear capability was adopted to replace the ReLu activation function of the original model. Since images in natural environments usually have complex background and foreground information, the robustness of HardSwish helped the model better handling these situations and further improving the detection accuracy. With the above improvements, the Camellia-YOLACT model was constructed and experimentally validated by testing the Camellia oleifera anthracnose leaf image dataset. [Results and Discussions] A transfer learning approach was used for experimental validation on the Camellia oleifera anthracnose severity grading dataset, and the results of the ablation experiments showed that the mAP₇₅ of Camellia-YOLACT proposed in this study was 86.8%, mAP_all was 78.3%, mAR was 91.6% which were 5.7%, 2.5% and 7.9% higher than YOLACT model. In the comparison experiments, Camellia-YOLACT performed better than Segmenting Objects by Locations (SOLO) in terms of both accuracy and speed, and its detection speed was doubled compared to Mask R-CNN algorithm. Therefore, the Camellia-YOLACT algorithm was suitable in Camellia oleifera gardens for anthracnose real-time segmentation. In order to verify the outdoors detection performance of Camellia-YOLACT model, 36 groups of Camellia oleifera anthracnose grading experiments were conducted. Experimental results showed that the grading correctness of Camellia oleifera anthracnose injection severity reached 94.4%, and the average absolute error of K-value was 1.09%. Therefore, the Camellia-YOLACT model proposed in this study has a better performance on the grading of the severity of Camellia oleifera anthracnose. [Conclusions] The Camellia-YOLACT model proposed got high accuracy in leaf and anthracnose segmentation of Camellia oleifera, on the basis of which it can realize automatic grading of the severity of Camellia oleifera anthracnose. This research could provide technical support for the precise control of Camellia oleifera diseases.

Key words: Camellia oleifera, leaf disease, anthracnose, BiFPN, YOLACT, Transformer, deep learning

聂刚刚, 饶洪辉, 李泽锋, 刘木华. 基于改进YOLACT的油茶叶片炭疽病感染严重程度分级模型[J]. 智慧农业(中英文), 2024, 6(3): 138-147.

NIE Ganggang, RAO Honghui, LI Zefeng, LIU Muhua. Severity Grading Model for Camellia Oleifera Anthracnose Infection Based on Improved YOLACT[J]. Smart Agriculture, 2024, 6(3): 138-147.

图/表 13

图1

图2

表1

图3

图4

图5

图6

图7

表2

图8

表3

图9

表4

参考文献 23

1	张立伟, 王辽卫. 我国油茶产业的发展现状与展望[J]. 中国油脂, 2021, 46(6): 6-9, 27.
	ZHANG L W, WANG L W. Prospect and development status of oil-tea camellia industry in China[J]. China oils and fats, 2021, 46(6): 6-9, 27.
2	吴鹏飞, 姚小华. 种植密度对普通油茶炭疽病病害发生的影响[J]. 中国油料作物学报, 2019, 41(3): 455-460.
	WU P F, YAO X H. Effect of planting density on anthracnose occurrence of Camellia oleifera [J]. Chinese journal of oil crop sciences, 2019, 41(3): 455-460.
3	张蕊, 李锦涛. 基于深度学习的场景分割算法研究综述[J]. 计算机研究与发展, 2020, 57(4): 859-875.
	ZHANG R, LI J T. A Survey on Algorithm Research of Scene Parsing Based on Deep Learning[J]. Journal of computer research and development, 2020, 57(4): 859-875.
4	HAQUE M A, MARWAHA S, ARORA A, et al. A lightweight convolutional neural network for recognition of severity stages of maydis leaf blight disease of maize[J]. Frontiers in plant science, 2022, 13: ID 1077568.
5	PRABHAKAR M, PURUSHOTHAMAN R, AWASTHI D P. Deep learning based assessment of disease severity for early blight in tomato crop[J]. Multimedia tools and applications, 2020, 79(39): 28773-28784.
6	TENDANG S, CHAMNONGTHAI K. Rice-disease severity level estimation using deep convolutional neural network[C]// 2021 36th International Technical Conference on Circuits/Systems, Computers and Communications (ITC-CSCC). Piscataway, New Jersey, USA: IEEE, 2021: 1-4.
7	万军杰, 祁力钧, 卢中奥, 等. 基于迁移学习的GoogLeNet果园病虫害识别与分级[J]. 中国农业大学学报, 2021, 26(11): 209-221.
	WAN J J, QI L J, LU Z A, et al. Recognition and grading of diseases and pests in orchard by GoogLeNet based on transfer learning[J]. Journal of China agricultural university, 2021, 26(11): 209-221.
8	LIU B, DING Z, TIAN L, et al. Grape leaf disease identification using improved deep convolutional neural networks[J]. Frontiers in plant science, 2020, 11: ID 1082.
9	王振, 张善文, 赵保平. 基于级联卷积神经网络的作物病害叶片分割[J]. 计算机工程与应用, 2020, 56(15): 242-250.
	WANG Z, ZHANG S W, ZHAO B P. Crop diseases leaf segmentation method based on cascade convolutional neural network[J]. Computer engineering and applications, 2020, 56(15): 242-250.
10	GARG K, BHUGRA S, LALL B. Automatic quantification of plant disease from field image data using deep learning[C]// 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). Piscataway, New Jersey, USA: IEEE, 2021: 1965-1972.
11	GONÇALVES JULIANO P, PINTO FRANCISCO A C, QUEIROZ DANIEL M, et al. Deep learning architectures for semantic segmentation and automatic estimation of severity of foliar symptoms caused by diseases or pests[J]. Biosystems engineering, 2021, 210: 129-142.
12	茹佳棋, 吴斌, 翁翔, 等. 基于改进UNet++模型的葡萄黑腐病病斑分割和病害程度分级[J]. 浙江农业学报, 2023, 35(11): 2720-2730.
	RU J Q, WU B, WENG X, et al. Disease spot segmentation and disease degree classification of grape black rot based on improved UNet++ model[J]. Acta agriculturae Zhejiangensis, 2023, 35(11): 2720-2730.
13	邓朝, 纪苗苗, 任永泰. 基于Mask R-CNN的马铃薯叶片晚疫病量化评价[J]. 扬州大学学报(农业与生命科学版), 2022, 43(1): 135-142.
	DENG Z, JI M M, REN Y T. Quantitative evaluation of potato late blight disease based on Mask R-CNN[J]. Journal of Yangzhou university (agricultural and life science edition), 2022, 43(1): 135-142.
14	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). Piscataway, New Jersey, USA: IEEE, 2021: 10012-10022.
15	TAN M X, PANG R M, LE Q V. EfficientDet: scalable and efficient object detection[C]// 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, New Jersey, USA: IEEE, 2020: 10781-10790.
16	安徽省市场监督管理局. 茶炭疽病测报调查与防治技术规程: DB34/T 3863—2021 [S].
17	SHU J H, NIAN F D, YU M H, et al. An improved mask R-CNN model for multiorgan segmentation[J]. Mathematical Problems in Engineering, 2020, 2020: 1-11.
18	ZHANG Z, HUANG S, LIU X, et al. Adversarial attacks on YOLACT instance segmentation[J]. Computers & Security, 2022, 116: ID 102682.
19	SHAFIQ M, GU Z. Deep residual learning for image recognition: A survey[J]. Applied sciences, 2022, 12(18): ID 8972.
20	杨毅, 桑庆兵. 多尺度特征自适应融合的轻量化织物瑕疵检测[J]. 计算机工程, 2022, 48(12): 288-295.
	YANG Y, SANG Q B. Lightweight-fabric defect detection based on adaptive fusion of multiscale features[J]. Computer engineering, 2022, 48(12): 288-295.
21	ZHU L, LEE F, CAI J, et al. An improved feature pyramid network for object detection[J]. Neurocomputing, 2022, 483: 127-139.
22	蓝金辉, 王迪, 申小盼. 卷积神经网络在视觉图像检测的研究进展[J]. 仪器仪表学报, 2020, 41(4): 167-182.
	LAN J H, WANG D, SHEN X P. Research progress on visual image detection based on convolutional neural network[J]. Chinese journal of scientific instrument, 2020, 41(4): 167-182.
23	WANG X, KONG T, SHEN C, et al. Solo: Segmenting objects by locations[C]// Computer Vision-ECCV 2020. ECCV 2020. Lecture Notes in Computer Science. Cham, Germany: Springer, 2020: 649-665.

严重程度分级	分级标准
0级	无病斑
1级	K≤25%
2级	25%< K≤50%
3级	50%< K≤75%
4级	K >75%

编号	主干网络	BiFPN	HardSwish	mAP₇₅/%	mAP_all/%	mAR/%
0	ResNet50	×	×	81.1	75.8	83.7
1	Swin-T	×	×	86.7	78.0	91.2
2	Swin-T	√	×	86.5	78.1	91.3
3	Swin-T	√	√	86.8	78.3	91.6

模型	mAP₇₅/%	mAP_all/%	mAR/%	td/ms
YOLACT	81.1	75.8	83.7	46.85
Mask R-CNN	90.4	81.5	92.3	168.16
SOLO	87.1	77.8	90.4	75.73
Camellia-YOLACT	86.8	78.3	91.6	53.50

编号	真实值		预测值		K绝对误差/%	编号	真实值		预测值		K绝对误差/%
编号	K/%	等级	K/%	等级	K绝对误差/%	编号	K/%	等级	K/%	等级	K绝对误差/%
1	58.31	3	57.28	3	1.03	19	80.68	4	77.11	4	3.57
2	43.14	2	43.05	2	0.09	20	23.92	1	21.31	1	2.61
3	0.00	0	0.00	0	0.00	21	6.26	1	6.24	1	0.02
4	7.32	1	6.84	1	0.48	22	16.55	1	13.69	1	2.86
5	26.08	2	24.56	1	1.52	23	22.32	1	20.28	1	2.04
6	8.54	1	8.03	1	0.51	24	7.83	1	7.04	1	0.79
7	63.88	3	63.03	3	0.85	25	21.80	1	21.50	1	0.30
8	10.39	1	9.34	1	1.05	26	0.00	0	0.00	0	0.00
9	52.65	3	49.12	2	3.53	27	8.18	1	7.67	1	0.51
10	76.44	4	75.34	4	1.10	28	12.97	1	10.95	1	2.02
11	53.34	3	53.26	3	0.08	29	19.50	1	18.38	1	1.12
12	0.00	0	0.00	0	0.00	30	13.77	1	13.55	1	0.22
13	0.00	0	0.00	0	0.00	31	28.68	2	27.64	2	1.04
14	12.19	1	11.29	1	0.90	32	7.21	1	4.50	1	2.71
15	18.18	1	17.99	1	0.20	33	14.81	1	13.07	1	1.74
16	47.29	2	46.20	2	1.09	34	5.04	1	4.72	1	0.32
17	62.31	3	59.92	3	2.39	35	11.06	1	10.50	1	0.56
18	36.39	2	35.86	2	0.53	36	18.22	1	16.85	1	1.37

[1]	李豪, 杜雨秋, 肖星竹, 陈彦羲. 基于深度学习的四川盆地丘陵区县域耕地遥感识别研究[J]. 智慧农业(中英文), 2024, 6(3): 34-45.
[2]	张荣华, 白雪, 樊江川. 复杂场景下害虫目标检测算法：YOLOv8-Extend[J]. 智慧农业(中英文), 2024, 6(2): 49-61.
[3]	张京, 赵泽瑄, 赵艳茹, 卜泓超, 吴星宇. 基于Bi-GRU和空-谱信息融合的油菜菌核病侵染区域高光谱图像分割方法[J]. 智慧农业(中英文), 2024, 6(2): 40-48.
[4]	庞春晖, 陈鹏, 夏懿, 章军, 王兵, 邹岩, 陈天娇, 康辰瑞, 梁栋. 用于小麦多生长阶段倒伏边界精准检测的分层交互特征金字塔网络[J]. 智慧农业(中英文), 2024, 6(2): 128-139.
[5]	张玉玉, 邴树营, 纪元浩, 严蓓蓓, 许金普. 基于改进YOLOv8s的玫瑰鲜切花分级方法[J]. 智慧农业(中英文), 2024, 6(2): 118-127.
[6]	张建华, 姚琼, 周国民, 吴雯迪, 修晓杰, 王健. 作物农艺性状与形态结构表型智能识别技术综述[J]. 智慧农业(中英文), 2024, 6(2): 14-27.
[7]	郭旺, 杨雨森, 吴华瑞, 朱华吉, 缪祎晟, 顾静秋. 农业大模型：关键技术、应用分析与发展方向[J]. 智慧农业(中英文), 2024, 6(2): 1-13.
[8]	王鹤榕, 陈英义, 柴莹倩, 徐玲, 于辉辉. 融合VoVNetv2和置换注意力机制的鱼群摄食图像分割方法[J]. 智慧农业(中英文), 2023, 5(4): 137-149.
[9]	李政凯, 于嘉辉, 潘时佳, 贾泽丰, 牛子杰. 冬季猕猴桃树单木骨架提取与冠层生长预测方法[J]. 智慧农业(中英文), 2023, 5(4): 92-104.
[10]	唐辉, 王铭, 于秋实, 张佳茜, 刘连涛, 王楠. 融合改进UNet和迁移学习的棉花根系图像分割方法[J]. 智慧农业(中英文), 2023, 5(3): 96-109.
[11]	泮玮婷, 孙梦丽, 员琰, 刘平. 基于深度学习ImCascade R-CNN的小麦籽粒表形鉴定方法[J]. 智慧农业(中英文), 2023, 5(3): 110-120.
[12]	管博伦, 张立平, 朱静波, 李闰枚, 孔娟娟, 汪焱, 董伟. 农业病虫害图像数据集构建关键问题及评价方法综述[J]. 智慧农业(中英文), 2023, 5(3): 17-34.
[13]	龙佳宁, 张昭, 刘晓航, 李云霞, 芮照钰, 余江帆, 张漫, FLORES Paulo, 韩哲雄, 胡灿, 王旭峰. 利用改进EfficientNetV2和无人机图像检测小麦倒伏类型[J]. 智慧农业(中英文), 2023, 5(3): 62-74.
[14]	张淦, 严海峰, 胡根生, 张东彦, 程涛, 潘正高, 许海峰, 沈书豪, 朱科宇. 基于深度学习语义分割和迁移学习策略的麦田倒伏面积识别方法[J]. 智慧农业(中英文), 2023, 5(3): 75-85.
[15]	刘易雪, 宋育阳, 崔萍, 房玉林, 苏宝峰. 基于无人机遥感和深度学习的葡萄卷叶病感染程度诊断方法[J]. 智慧农业(中英文), 2023, 5(3): 49-61.