基于深度学习的马铃薯花粉活力快速检测

doi:10.13560/j.cnki.biotech.bull.1985.2024-0511

摘要/Abstract

摘要：

【目的】传统马铃薯花粉活力检测方法依靠肉眼计数，存在效率低、准确性差等问题。本研究基于PaddlePaddle深度学习框架，通过比较不同模型，提出一种快速检测花粉活力的方法。【方法】首先收集花粉，使用2,3,5-氯化三苯基四氮唑（2,3,5-triphenyltetrazolium chloride，TTC）染色，通过显微镜拍照获取图像；利用Photoshop（PS）进行数据标注，分别标注有活力花粉和所有花粉，并将标签图转为单通道图像；选用SegFormer、U-Net和DeepLabV3模型进行训练，分割有活力花粉和所有花粉；最后使用Python OpenCV程序计数，计算花粉活力。【结果】与其他模型相比，SegFormer在两类数据集中的各项评估指标均为最优。相比于人工识别，OpenCV程序可以实现快速、批量计数，且误差小。【结论】通过图像处理技术可以快速、准确检测马铃薯花粉活力，进一步利用该方法快速鉴定了200份F₂株系的花粉活力，为马铃薯花粉活力表型采集奠定了基础。

关键词: 马铃薯, 花粉活力, 图像处理, 深度学习, SegFormer, OpenCV

Abstract:

【Objective】Traditional methods for detecting potato pollen viability rely on visual counting, which can be inefficient and inaccurate. In this study, a method for quickly detecting pollen viability was proposed based on PaddlePaddle deep learning framework by comparing different models.【Method】First, the pollens were stained with 2,3,5-triphenyltetrazolium chloride（TTC）and imaged using a microscope. The images were annotated by Photoshop（PS）. Viable and total pollens were labeled respectively, then the label images were converted into single-channel images. Three models, SegFormer, U-Net and DeepLabV3, were used for training to distinguish viable pollens and total pollens. Finally, a Python OpenCV program was used to count the pollen number and calculate pollen viability. 【Result】Compared with other models, SegFormer demonstrated the best performance in various evaluation indexes of the two datasets. Compared with manual recognition, the OpenCV program enabled fast and batch counting with less error.【Conclusion】Potato pollen viability can be detected quickly and accurately by image processing technology. This method was used to quickly identify the pollen viability of 200 F₂ individuals, providing a soild foundation for the collection of pollen viability in potato.

Key words: potato, pollen viability, image processing, deep learning, SegFormer, OpenCV

夏士轩, 耿泽栋, 祝光涛, 张春芝, 李大伟. 基于深度学习的马铃薯花粉活力快速检测[J]. 生物技术通报, 2024, 40(9): 123-130.

XIA Shi-xuan, GENG Ze-dong, ZHU Guang-tao, ZHANG Chun-zhi, LI Da-wei. Quick Detection of Potato Pollen Viability Based on Deep Learning[J]. Biotechnology Bulletin, 2024, 40(9): 123-130.

图/表 6

图1 花粉活力鉴定流程

Fig. 1 Identifying process of pollen viability

表1 不同模型对比实验的结果

Table 1 Comparative results of different algorithms

数据集 Dataset	模型 Model	交并比 IoU	精确率 Precision	召回率 Recall	F1分数 F1 score
有活力花粉	SegFormer	35.25	85.95	38.26	51.11
	U-Net	34.34	78.62	38.28	49.92
	DeepLabV3	31.34	78.19	35.04	47.35
所有花粉	SegFormer	75.65	82.67	88.64	84.78
	U-Net	58.50	74.56	72.58	73.08
	DeepLabV3	59.08	78.26	69.97	73.70

图2 不同尺度图片的分割效果原图为花粉TTC染色图像；伪彩色图为模型输出的花粉分割图像；掩膜图是为方便观察分割效果，将伪彩色图与原图掩膜后的结果；200 μm（左图）；400 μm（右图）

Fig. 2 Segmentation results of different-scale pictures The original image is the pollen stained by TTC; the pseudo-color image is the pollen segmentation image output by the model; the mask image is the result of masking the pseudo-color image and the original image for the convenience of observing the segmentation effect; 200 μm（left）; 400 μm（right）

图3 OpenCV程序识别效果原图为花粉TTC染色图像；掩膜图为模型输出的伪彩色图与原图掩膜的结果；识别图为OpenCV程序对识别到的花粉进行标记后输出的图像；比例尺200 μm

Fig. 3 Recognition effect by OpenCV program The original image is the pollen stained by TTC; the mask image is the result of masking the pseudo-color image output by the model with the original image; the recognition image is the output by the OpenCV program after labeling the recognized pollens; bar = 200 μm

表2 不同识别方法的结果对比

Table 2 Comparison of different identification methods

编号 No.	有活力花粉数The number of viable pollens			总花粉数The number of total pollens			花粉活力Pollen viability
编号 No.	程序计算值Calculated value by program	人工计算值Manually calculated value	误差 Error	程序计算值Calculated value by program	人工计算值Manually calculated value	误差 Error	程序计算值Calculated value by program/%	人工计算值Manually calculated value/%	误差 Error/%
2×-1	569	535	-34	1 074	1 021	-53	53	52	-1
2×-2	241	240	-1	442	457	15	55	53	-2
2×-3	297	296	-1	390	403	13	76	73	-3
2×-4	447	445	-2	528	553	25	84	80	-4
2×-5	355	357	2	483	511	28	73	70	-3
2×-6	158	154	-4	315	303	-12	50	51	1
2×-7	99	103	4	138	153	15	72	68	-4
2×-8	167	161	-6	235	238	3	71	68	-3
2×-9	142	139	-3	177	183	6	80	76	-4
2×-10	314	314	0	553	537	-16	57	58	1
4×-1	117	125	8	249	270	21	47	46	-1
4×-2	118	124	6	244	263	19	48	47	-1
4×-3	80	82	2	174	180	6	46	46	0
4×-4	71	70	-1	187	199	12	38	35	-3
4×-5	72	79	7	240	246	6	30	32	2
4×-6	73	81	8	245	257	12	30	32	2
4×-7	70	74	4	204	218	14	34	34	0
4×-8	75	73	-2	188	206	18	40	35	-4
4×-9	69	75	6	232	251	19	30	30	0
4×-10	56	60	4	186	198	12	30	30	0

图4 200份马铃薯F2株系花粉活力统计

Fig. 4 Pollen viability statistics of 200 potato F2 individuals

参考文献 32

[1]	Hardigan MA, Laimbeer FPE, Newton L, et al. Genome diversity of tuber-bearing Solanum uncovers complex evolutionary history and targets of domestication in the cultivated potato[J]. Proc Natl Acad Sci USA, 2017, 114(46): E9999-E10008.
[2]	Camire ME, Kubow S, Donnelly DJ. Potatoes and human health[J]. Crit Rev Food Sci Nutr, 2009, 49(10): 823-840. doi: 10.1080/10408390903041996 pmid: 19960391
[3]	Monro J, Mishra S, Blandford E, et al. Potato genotype differences in nutritionally distinct starch fractions after cooking, and cooking plus storing cool[J]. J Food Compos Anal, 2009, 22(6): 539-545.
[4]	Aliche EB, Theeuwen TPJM, Oortwijn M, et al. Carbon partitioning mechanisms in potato under drought stress[J]. Plant Physiol Biochem, 2020, 146: 211-219.
[5]	Simakov EA, Anisimov BV, Yashina IM, et al. Potato breeding and seed production system development in Russia[J]. Potato Res, 2008, 51(3): 313-326.
[6]	Celis C, Scurrah M, Cowgill S, et al. Environmental biosafety and transgenic potato in a centre of diversity for this crop[J]. Nature, 2004, 432(7014): 222-225.
[7]	Gebhardt C, Valkonen JP. Organization of genes controlling disease resistance in the potato genome[J]. Annu Rev Phytopathol, 2001, 39: 79-102. pmid: 11701860
[8]	Lindhout P, Meijer D, Schotte T, et al. Towards F1 hybrid seed potato breeding[J]. Potato Res, 2011, 54(4): 301-312.
[9]	Boavida LC, Vieira AM, Becker JD, et al. Gametophyte interaction and sexual reproduction: how plants make a zygote[J]. Int J Dev Biol, 2005, 49(5/6): 615-632.
[10]	Dafni A, Firmage D. Pollen viability and longevity: practical, ecological and evolutionary implications[J]. Plant Syst Evol, 2000, 222(1): 113-132.
[11]	Chen DJ, Neumann K, Friedel S, et al. Dissecting the phenotypic components of crop plant growth and drought responses based on high-throughput image analysis[J]. Plant Cell, 2014, 26(12): 4636-4655.
[12]	Walter A, Liebisch F, Hund A. Plant phenotyping: from bean weighing to image analysis[J]. Plant Methods, 2015, 11: 14. doi: 10.1186/s13007-015-0056-8 pmid: 25767559
[13]	Song P, Wang JL, Guo XY, et al. High-throughput phenotyping: breaking through the bottleneck in future crop breeding[J]. Crop J, 2021, 9(3): 633-645. doi: 10.1016/j.cj.2021.03.015
[14]	García-Fortea E, García-Pérez A, Gimeno-Páez E, et al. A deep learning-based system(microscan)for the identification of pollen development stages and its application to obtaining doubled haploid lines in eggplant[J]. Biology, 2020, 9(9): 272.
[15]	周成全, 叶宏宝, 俞国红, 等. 基于机器视觉与深度学习的西兰花表型快速提取方法研究[J]. 智慧农业, 2020, 2(1): 121-132.
	Zhou CQ, Ye HB, Yu GH, et al. A fast extraction method of broccoli phenotype based on machine vision and deep learning[J]. Smart Agric, 2020, 2(1): 121-132. doi: 10.12133/j.smartag.2020.2.1.201912-SA003
[16]	赵越, 赵辉, 姜永成, 等. 基于深度学习的马铃薯叶片病害检测方法[J]. 中国农机化学报, 2022, 43(10): 183-189.
	Zhao Y, Zhao H, Jiang YC, et al. Detection method of potato leaf diseases based on deep learning[J]. J Chin Agric Mech, 2022, 43(10): 183-189.
[17]	胡松涛, 翟瑞芳, 王应华, 等. 基于多源数据的马铃薯植株表型参数提取[J]. 智慧农业, 2023, 5(1):132-145.
	Hu ST, Zhai RF, Wang YH, et al. Extraction of potato plant phenotypic parameters based on multi-source data[J]. Smart Agric, 2023, 5(1): 132-145. doi: 10.12133/j.smartag.SA202302009
[18]	Gao Y, Li YL, Jiang RB, et al. Enhancing green fraction estimation in rice and wheat crops: a self-supervised deep learning semantic segmentation approach[J]. Plant Phenomics, 2023, 5: 0064.
[19]	Tello J, Montemayor MI, Forneck A, et al. A new image-based tool for the high throughput phenotyping of pollen viability: evaluation of inter- and intra-cultivar diversity in grapevine[J]. Plant Methods, 2018, 14: 3. doi: 10.1186/s13007-017-0267-2 pmid: 29339970
[20]	Tan ZH, Yang J, Li QY, et al. PollenDetect: an open-source pollen viability status recognition system based on deep learning neural networks[J]. Int J Mol Sci, 2022, 23(21): 13469.
[21]	马艳军, 于佃海, 吴甜, 等. 飞桨:源于产业实践的开源深度学习平台[J]. 数据与计算发展前沿, 2019, 1(5):105-115.
	Ma YJ, Yu DH, Wu T, et al. PaddlePaddle: an open-source deep learning platform from industrial practice[J]. Frontiers of Data and Computing, 2019, 1(5): 105-115.
[22]	Hao SJ, Zhou Y, Guo YR. A brief survey on semantic segmentation with deep learning[J]. Neurocomputing, 2020, 406: 302-321.
[23]	Zhang CZ, Yang ZM, Tang D, et al. Genome design of hybrid potato[J]. Cell, 2021, 184(15): 3873-3883.e12. doi: 10.1016/j.cell.2021.06.006 pmid: 34171306
[24]	Sun C, Shrivastava A, Singh S, et al. Revisiting unreasonable effectiveness of data in deep learning era[C]// 2017 IEEE International Conference on Computer Vision(ICCV). Venice, Italy. Piscataway, NJ: IEEE, 2017: 843-852.
[25]	Shorten C, Khoshgoftaar TM. A survey on image data augmentation for deep learning[J]. J Big Data, 2019, 6(1): 60.
[26]	Su D, Kong H, Qiao YL, et al. Data augmentation for deep learning based semantic segmentation and crop-weed classification in agricultural robotics[J]. Comput Electron Agric, 2021, 190: 106418.
[27]	Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: transformers for image recognition at scale[J]. arXiv E Prints, 2020: arXiv: 2010.11929.
[28]	Xie EZ, Wang WH, Yu ZD, et al. SegFormer: simple and efficient design for semantic segmentation with transformers[EB/OL]. 2021: arXiv: 2105. 15203. http://arxiv.org/abs/2105.15203.
[29]	Ronneberger O, Fischer P, Brox T. U-net: convolutional networks for biomedical image segmentation[C]// International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer, 2015: 234-241.
[30]	Chen LC, Papandreou G, Schroff F, et al. Rethinking atrous convolution for semantic image segmentation[EB/OL]. 2017: arXiv: 1706.05587. http://arxiv.org/abs/1706.05587.
[31]	Yang WN, Feng H, Zhang XH, et al. Crop phenomics and high-throughput phenotyping: past decades, current challenges, and future perspectives[J]. Mol Plant, 2020, 13(2): 187-214. doi: S1674-2052(20)30008-3 pmid: 31981735
[32]	仇瑞承, 魏爽, 张漫, 等. 作物表型组学测量方法综述[J]. 中国农业文摘-农业工程, 2019, 31(1): 23-36, 55.
	Qiu RC, Wei S, Zhang M, et al. Summary of crop phenotypic omics measurement methods[J]. Agric Sci Eng China, 2019, 31(1): 23-36, 55.