基于深度学习的近红外光谱图像处理的农作物虫害检测研究  

Research on crop pest detection based on near-infrared spectroscopy image processing based on deep learning

作　　者：陆娇娇[1,2] 肖衡 LU Jiajiao;XIAO heng(Sanya University School of Information&Intelligence Engineering,Sanya Hainan 572000,China;Academician Workstation of Guoliang Chen,Sanya Hainan 572000,China)

机构地区：[1]三亚学院信息与智能工程学院,海南三亚572000 [2]陈国良院士工作站,海南三亚572000

出　　处：《激光杂志》2023年第8期221-225,共5页Laser Journal

基　　金：海南省自然科学基金(No.621QN0900);三亚市科技工业信息化局(No:2021GXYL58)。

摘　　要：采集农作物虫害图像时常因外界干扰使特征区域划分不明,导致虫害检测精度不足,为此,提出基于深度学习的农作物虫害近红外光谱图像分割算法。利用分块匹配模型对近红外光谱图像进行视觉纹理提取,通过高斯滤波技术去除噪声干扰。采用边缘特征重构和角点匹配方法,构建近红外光谱图像分割和相似度度量模型,采用分块帧点检测的方法对目标关键特征点定位和包络检测,多模态特征匹配后得到虫害特征分布,完成虫害检测。测试结果表明,该方法对虫害定位均方根误差平均为0.14,耗时平均为256.4 ms,检测精度平均为96.68%,具有较好的应用性能。When collecting images of crop pests,the characteristic areas are often unclear due to external interfer-ence,which leads to insufficient accuracy of pest detection.Therefore,this paper proposes a near-infrared spectral image segmentation algorithm of crop pests based on deep learning.Block matching model is used to extract the visual texture of near-infrared spectral images,and Gaussian filtering technology is used to remove noise interference.The edge feature reconstruction and corner matching methods are used to construct the near-infrared spectrum image seg-mentation and similarity measurement model,and the block frame point detection method is used to locate and detect the key feature points of the target.After multi-modal feature matching,the pest feature distribution is obtained,and the pest detection is completed.The test results show that the average root mean square error of the method is 0.14,the average time-consuming is 256.4 ms,and the average detection accuracy is 96.68%.It has good application performance.

关 键 词：深度学习 近红外光谱图像 农作物 虫害检测 角点匹配 包络检测 

分 类 号：TN247[电子电信—物理电子学]