中国农业科学 ›› 2018, Vol. 51 ›› Issue (11): 2084-2093.doi: 10.3864/j.issn.0578-1752.2018.11.006

• 耕作栽培·生理生化·农业信息技术 • 上一篇    下一篇

基于多特征融合和稀疏表示的农业害虫图像识别方法

张永玲1,姜梦洲1,俞佩仕1,姚青1,杨保军2,唐健2

 
  

  1. 1浙江理工大学信息学院,杭州 3100182中国水稻研究所/水稻生物重点实验室,杭州310006
  • 收稿日期:2017-10-30 出版日期:2018-06-01 发布日期:2018-06-01
  • 通讯作者: 姚青,E-mail:q-yao@126.com。唐健,E-mail:tangjian@caas.net
  • 作者简介:张永玲,E-mail:zyling7795@126.com
  • 基金资助:
    国家863计划资助项目(2013AA102402)、浙江省公益技术研究计划项目(LGN18C140007)

Agricultural Pest Identification Based on Multi-Feature Fusion and Sparse Representation

ZHANG YongLing1, JIANG MengZhou1, YU PeiShi1, YAO Qing1, YANG BaoJun2, TANG Jian2   

  1. 1School of Information and technology, Zhejiang Sci-Tech University, Hangzhou 310018; 2China National Rice Research Institute/State Key Laboratory of Rice Biology, Hangzhou 310006
  • Received:2017-10-30 Online:2018-06-01 Published:2018-06-01

摘要: 【目的】在农业害虫测报中,常常需要从大量的昆虫中识别出几种重要的测报害虫。目前基于图像的农业害虫识别研究,大部分是在有限种类有限样本量基础上进行的农业害虫识别。本研究为了从大量的水稻昆虫图像中识别出9种水稻测报害虫,尝试提出了一种基于多特征融合和稀疏表示的农业害虫图像识别方法。【方法】首先,为了获得最优的农业害虫识别模型,将所有图像进行旋转使昆虫头朝上,按照1﹕2长宽比裁剪图像,使昆虫居中并占据图像大部分区域,将图像进行等比例缩放至统一尺寸48×96像素。提取所有昆虫的HSV颜色特征、局部特征中的HOG特征、Gabor特征和LBP特征。然后,利用单一特征和融合特征分别对训练样本构建过完备字典,字典中的每一个列向量表示一个训练样本,且满足同一类训练样本均在同一个子空间中;应用过完备字典对测试图像进行多特征稀疏表示,通过求解l1范数意义下的优化问题获取稀疏解,使得除测试样本所在的类别外其他的训练样本的系数都是零或接近零的数值。最后,计算稀疏集中指数阈值,用于判断测试样本的有效性,如果测试样本的稀疏集中指数大于该阈值,则认为最小残差所对应的类别即为测试样本的类别,否则认为该测试样本为非测报昆虫。同时,利用相同的特征和训练样本训练SVM分类器对测试样本进行测试,与稀疏表示害虫识别模型进行比较。【结果】利用单一特征训练的稀疏表示害虫识别模型中,基于HOG特征的稀疏表示识别模型获得了9种测报害虫较高的识别率和较低的误检率,分别为87.0%和7.5%;利用颜色特征分别与3种局部特征进行结合获得的稀疏表示识别模型,测试结果表明,基于颜色和HOG特征的稀疏表示识别模型获得了最高的识别率和最低的误检率,分别为90.1%和5.2%;将颜色、HOG和Gabor 3个特征结合获得的稀疏表示识别模型,识别率下降为83.5%,误检率上升为10.3%。利用同样的特征或特征融合训练得到的支持向量机分类器,识别率均低于对应特征获得的稀疏表示识别模型的识别率,而误检率均高于对应特征训练的稀疏表示害虫识别模型的误检率。【结论】基于颜色和HOG 融合特征的稀疏表示识别模型获得了较高的农业害虫识别率和较低的误检率;通过稀疏集中指数阈值,有效地排除了非测报昆虫,实现了从大量的农业昆虫中自动识别出需要测报的害虫。

关键词: 农业测报害虫, 特征融合, 稀疏表示, 识别模型, 支持向量机

Abstract: 【Objective】In agricultural pest forecasting, it is often necessary to identify several important pests from a large number of insects. At present, most of the researches on agricultural pest identification are based on limited pest species and limited sample sizes. In order to identify nine species of rice forecasting pests from a large number of agricultural insect images, a method based on multi-feature and sparse representation for pest image identification was proposed in this paper. 【Method】 Firstly, in order to obtain an optimal identification model of agricultural pests, all pest images were rotated to make insect head upright,  cropped with 1:2 aspect ratio to make insect center and take up a large portion of the image, and scaled to a uniform size with 48×96 pixels. The HSV color features, LBP features, Gabor features and HOG (histogram of oriented gradient) features of each image were extracted. Then, the overcomplete dictionaries based on single features or multi-feature were constructed and each column vector represented a training sample. The same training sample species were in the same subspace. a testing sample was sparsely represented by an overcomplete dictionary and a sparse solution was obtained by solving the optimization of the l1 norm to make different training sample species with zero or near-zero coefficient. Finally, the threshold value of sparse concentration index was used to determine the validity of the testing sample. If the sparse concentration index of a testing sample was greater than the threshold value, the pest was identified as the species with the minimizing reconstruction error. Otherwise, the testing sample was judged to be a non-forecasting pest. The support vector machine (SVM) classifiers were trained on the same features and training samples to compare with the sparse representation identification model of agricultural pests. 【Result】 In the sparse representation identification models trained on single feature, the model based on HOG features could get the higher identification rate of 87.0% and lower false detection rate of 7.5% in the nine rice forecasting pests. The sparse representation identification model based on color and HOG features obtained the highest identification rate of 90.1% and lowest false detection rate of 5.2%. The identification rate decreased and the false detection rate rose in the sparse representation models based on color, HOG and Gabor features. The identification rates of the SVM classifiers were lower and the false detection rates were higher than the sparse representation models trained on the same features. 【Conclusion】 The sparse representation pest identification model based on color and HOG features obtained the higher identification rate and lower false detection rate of agricultural forecasting pests. The threshold value of sparse concentration index could effectively exclude the non-forecasting insects. The sparse representation pest identification model could automatically identify the forecasting pests from a large number of insects.

Key words: agricultural forecasting pest, feature fusion, sparse representation, identification model, support vector machine