铁质破片的高光谱图像识别研究  

Research on Hyperspectral Image Recognition of Iron Fragments

作　　者：郭峰 赵冬娥[1] 杨学峰[1] 褚文博 张斌[1] 张大舜 孟凡军 GUO Feng;ZHAO Dong-e;YANG Xue-feng;CHU Wen-bo;ZHANG Bin;ZHANG Da-shun;MENG Fan-jun(School of Information and Communication Engineering,North University of China,Taiyuan 030051,China;School of Mechanical and Electrical Engineering,North University of China,Taiyuan 030051,China;The Fifth Research Institute of China North Industries Group,Changchun 130012,China)

机构地区：[1]中北大学信息与通信工程学院,山西太原030051 [2]中北大学机电工程学院,山西太原030051 [3]中国兵器工业集团第五研究所,吉林长春130012

出　　处：《光谱学与光谱分析》2023年第4期997-1003,共7页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(62205307);中国兵器工业集团第五研究所创新基金项目资助。

摘　　要：高光谱成像凭借高的光谱分辨率、图谱合一、波段多的特点,能够为待分类目标提供多维的参考信息,从而提高分类精度。爆炸破片的识别回收能够为爆炸威力的评估和防爆措施的设计提供参考。针对当前破片检测中多采用可见光波段或红外波段等单个波段进行检测,忽略了破片目标与背景对不同波长的光有着不同的吸收程度,没有将多波段破片特征充分利用,为此结合高光谱检测手段,提出了一种空间分割结合光谱信息的爆炸破片识别方法。在实验室环境下,首先采集铁质破片、石头、树叶的高光谱图像,对采集的样本图像数据做预处理,包括去噪声以及黑白校正反演反射率信息等,感兴趣区域随机提取三类样本像素点共750个,随机选取600个点作为训练集其余作为测试集,通过训练后得到预测准确度分别为88%、88%、94%的决策树模型。其次模拟了铁质破片散落在含有石头树叶的沙土中的场景并采集其高光谱数据,通过前后级联的空谱融合方法,在空域经过图像增强和去噪等预处理之后,采用边缘检测结合区域生长以及形态学处理的方法对空间图像进行分割,得到沙土上有形态的目标,空间分割的交并比(IOU)达到93.5%,真阳率(TPR)达到97.4%;然后结合光谱域训练得到的决策树模型,对各个分割区域的每个像素点进行谱域的类型识别,参与分类的三类像素点个数分别为146172、50484、213438,识别准确度分别为87%、86%、96%;最后将分类结果可视化,以每个区域像素点最多的一类代表该区域类别,将目标破片与石子和树叶两种背景进行了准确的识别,以标定后的分割图像为标准,三类像素点个数分别为155502、52045、217794,识别率分别为94%、97%、98%。分析结果表明空间分割结合光谱信息的识别方法能够有效利用空间和高光谱的特征信息对铁质破片目标进行准确识别。同时验证了使用�Hyperspectral imaging can provide multi-dimensional reference information for the target to be classified by its high spectral resolution,spectrum integration and multiple bands,thus improving the classification accuracy.The identification and recovery of explosive fragments can provide a reference for evaluating explosive power and designing explosion-proof measures.Because of the current fragment detection,single bands such as visible light band or infrared band are mostly used for detection,ignoring that the fragment target and background have different degrees of absorption of light of different wavelengths and do not make full use of the characteristics of multi-band fragments.Therefore,this paper combines hyperspectral detection means and proposes a method of explosive fragment recognition based on spatial segmentation and spectral information.In the laboratory environment,first collect the hyperspectral images of iron fragments,rocks and leaves,and preprocess the collected sample image data,including noise removal and black and white correction to retrieve the reflectivity information.Randomly extract 750 sample pixels of three types from the region of interest,and randomly select 600points as the rest of the training set as the test set.After training,a decision tree model with a prediction accuracy of 88%,88%and 94%is obtained.Secondly,the scene of iron fragments scattered in the sand with stone leaves is simulated,and its hyperspectral data is collected.Through the cascade space spectrum fusion method,after image enhancement and denoising in the spatial domain,the spatial image is segmented using edge detection combined with region growth and morphological processing methods to obtain the morphological targets on the sand.The intersection and union ratio(IOU)of spatial segmentation reaches 93.5%,and the true positive rate(TPR)reached 97.4%;Then,combined with the decision tree model trained in the spectral domain,each pixel point in each segmentation area is identified in the spectral domain.The number t

关 键 词：高光谱成像 空间分割 决策树 铁质破片 分类识别 

分 类 号：O433.4[机械工程—光学工程]