作者：吴建¹, 王红军²

作者单位：1. 广安职业技术学院，四川 广安 638000;
2. 西南交通大学，四川 成都 611756

关键词：玉米干旱检测;计算机视觉;CNN;Fisher向量

摘要：

近年来，随着计算机视觉技术的快速发展，越来越多基于图像的应用被开发出来用以解决农业生产中的问题。目前对农作物生长和产量影响最大的就是植物干旱胁迫问题。传统卷积神经网（CNN）使用直接回归玉米叶片数的算法，在实际实验中误差较大，因此该文引入CNN和Fisher向量编码联合提取多尺度特征的方法。在设计特征提取网络时参考GoogLeNet中多尺度卷积核结构，使其更适合于提取不同尺寸叶片的特征。然后，利用Fisher向量对部分中间层特征图编码，增强特征的表达能力。最后，使用随机森林回归叶片数量，并在实验中展示不同干旱程度对玉米叶片数量的影响：在一定条件下，土壤含水量与叶片数量呈正相关。该文方法与现有算法进行对比，叶片平均误差以及均方误差分别减少0.011与0.382，具有一定的研究价值。

Drought detection of crop plants based on convolutional neural network
WU Jian¹, WANG Hongjun²
1. Guang’an Vocational and Technology College, Guang’an 638000, China;
2. Southwest Jiaotong University, Chengdu 611756, China
Abstract: In recent years, with the rapid development of computer vision technology, more and more image-based applications have been developed to solve the problems in agricultural production. At present, plant drought stress is the most important factor affecting crop growth and yield. The traditional convolutional neural network uses the direct regression algorithm to count maize leaf number, which has a large error in the actual experiment. Therefore, this paper introduces the method of CNN and Fisher vector coding to extract multi-scale features. In the design of feature extraction network, we refer to the multi-scale convolution core structure of Google LeNet to make it more suitable for extracting the characteristics of blades of different sizes. Then, Fisher vectors are used to encode the middle-level feature maps to enhance the ability of feature expression. Finally, the number of leaves was regressed by random forest, and the effects of different drought degrees on the number of maize leaves were demonstrated in the experiment. Under certain conditions, The soil moisture content is positively correlated with the number of leaves. The method in this paper is compared with the existing algorithms, and the average error and mean square error of the leaves are reduced by 0.011 and 0.382 respectively, which has certain research value.
Keywords: maize drought detection;computer vision;CNN;Fisher vector
2022, 48(4):102-109  收稿日期: 2021-03-26;收到修改稿日期: 2021-04-25
基金项目: 国家自然科学基金（61773324，61603313）
作者简介: 吴建（1972-），男，四川广安市人，副教授，主要从事机器视觉与机器学习方面的研究
参考文献
[1] 刘桂雄, 张瑜, 蔡柳依婷. 机器视觉检测图像拼接融合技术研究进展[J]. 中国测试, 2020, 46(1): 1-6
[2] 陆昊. 玉米雄穗发育状态自动观测中的图像理解与分析[D]. 武汉：华中科技大学, 2018.
[3] 韩伟聪, 鲍光海. 基于机器视觉的竹材尺寸测量系统设计[J]. 中国测试, 2016, 42(7): 74-78
[4] 徐贵力, 毛罕平, 胡永光. 基于计算机视觉技术参考物法测量叶片面积[J]. 农业工程学报, 2002, 18(1): 154-157
[5] 朱军伟, 张彦娥, 王晓敏, 等. 基于图像纹理的玉米叶片叶绿素含量分析研究[J]. 微计算机信息, 2010, 26(10): 106-108
[6] 孙俊, 谭文军, 毛罕平, 等. 基于改进卷积神经网络的多种植物叶片病害识别[J]. 农业工程学报, 2017, 33(19): 209-215
[7] 张彦娥, 李民赞, 张喜杰, 等. 基于计算机视觉技术的温室黄瓜叶片营养信息检测[J]. 农业工程学报, 2005(8): 102-105
[8] PALOSCIA S, PAMPALONII P. Microwave remote sensing of plant water stress[J]. Remote Sensing of Environment, 1984, 16(3): 249-255
[9] JACKSON R D, IDSO S B, REGINATO R J, et al. Canopy temperature as a crop water stress indicator[J]. Water Resources Research, 1981, 17(4): 1133-1138
[10] NI Z, LIU Z, HUO H, et al. Early water stress detection using leaf-level measurements of chlorophyll fluorescence and temperature data[J]. Remote Sensing, 2015, 7(3): 3232-3249
[11] WALTER A, SCHURR U. The modular character of growth in Nicotiana tabacum plants under steady-state nutrition[J]. Journal of Experimental Botany, 1999, 50(336): 1169-1177
[12] AICH S, STAVNESS I . Leaf counting with deep convolutional and deconvolutional networks[C]// Proceedings of the IEEE International Conference on Computer Vision. 2017: 2080-2089.
[13] SIMONYAN K, PARKHI O M, VEDALDI A, et al. Fisher vector faces in the wild[C]//British Machine Vision Conference. 2013.
[14] PERRONNIN F, DANCE C. Fisher kernels on visual vocabularies for image categorization[C]// 2007 IEEE Conference on Computer Vision and Pattern Recognition. IEEE Computer Society, 2007.
[15] VINYALS O, JIA Y, DENG L, et al. Learning with recursive perceptual representations[C]//Advances in Neural Information Processing Systems. 2012: 2825-2833.
[16] KRIZHEVSKY A, SUTSKEVER I, HINTON G E. Imagenet classification with deep convolutional neural networks[C]//Advances in Neural Information Processing Systems. 2012: 1097-1105.
[17] SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[J]. Computer Science, 2014, 9(1): 135-146