高光谱成像的垃圾分类识别研究
|
摘要
高光谱成像因光谱分辨率高、图谱合一、可实现快速无损检测等特点现已广泛应用于农业、医学、遥感等领域。现有的对可回收生活垃圾检测与分类的方法,都存在检测时间长,分类效率低,而大量多种垃圾无法同时快速分拣等问题。考虑到不同类别的生活垃圾由于其主要组成分子结构的差异,对不同波长的光有不同的吸收特性。高光谱图像在记录待分类垃圾的空间信息的同时,可以获得垃圾对不同波长的光的反射率光谱信息,通过建立识别分类模型对反射率光谱信息进行分析可以实现对高光谱图像中待分类垃圾的识别与分类。收集常见纸质、塑料、木质三种材料的可回收的垃圾样本,包括塑料瓶、食品包装袋、塑料玩具(饰品)碎片、一次性筷子、雪糕棒、木制家具碎片、木制包装盒、废旧课本、广告纸、办公用纸等多种物品共30个样本,进行清洗和裁剪处理,避免样本表面污渍对样本反射率产生影响。利用高光谱成像系统采集样本在近红外(780～1 000 nm)范围内的高光谱图像,其中18个样本做训练样本集, 12个样本做测试样本集。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处理;通过主成分分析(PCA)方法对训练样本集感兴趣区域(ROI)进行分析,提取到的特征波段为795.815, 836.869, 885.619, 916.409, 929.239, 934.37, 957.463, 972.858和988.253 nm;在特征波段下分别提取这三种类别垃圾的参考光谱,通过光谱角度填图法(SAM)对测试样本ROI区域内提取的测试样本点集在特征波段下与参考光谱进行匹配,由匹配程度进行样本点归类,分析结果表明,测试样本集中纸制样本(A类别)、塑料样本(B类别)、木制样本(C类别)的分类准确度分别为100%, 98%和100%,测试样本点集整体的分类准确度为99.33%;通过Fisher判别方法分析训练样本集得出判别函数式和判别准则,对测试样本点集分类,评价结果为A, B和C类样本分类准确度分别为100%, 100%和97%,测试样本点集整体分类准确度为99%。通过SAM和Fisher两种判别方法对测试样本集的光谱图像进行目标物的检测与分类,结果表明,利用SAM判别方法在可回收垃圾的高光谱图像中实现检测与分类有更高的分类准确度,可达到99.33%。同时,也验证了使用高光谱成像进行可回收垃圾快速分类的科学性以及可行性,对未来系统化、机械化、智能化地解决生活中可回收垃圾的分类具有一定的实用意义。
Hyperspectral imaging technology is profoundly applied into the fields of agriculture, medicine and remote sensing due to its high spectral resolution, merged image-spectrum, and fast non-destructive testing. While the method used now has the defects of long-term testing period, poor efficiency and sorting asynchrony. Spectral image can identify and classify the target garbage by establishing a recognition and classification model and analyzing reflectance spectrum information based on the facts that different materials of domestic garbage, due to their different molecular structures, will absorb different wavelengths of light and the hyperspectral image can obtain the spatial information and the reflectance spectral information from different-wavelength illumination of the target garbage. Collected recyclable garbage samples of common paper, plastic and wood materials, including plastic bottles, food packaging bags, plastic toys(jewelry) pieces, disposable chopsticks, ice cream bars, wooden furniture pieces, wooden boxes, waste textbooks, advertising paper, office paper and other items, 30 in total. And cleaned and cut them to avoid the influence of sample surface stains on the sample reflectivity. Hyperspectral imaging systems were used to acquire hyperspectral images of the sample in the near-infrared(780 ～1 000 nm) formed 18 training samples and 12 test samples. Pre-processed the collected sample image by de-noising and black-and-white correction inversion of reflectivity information. Then analyzed the region of interest of training samples by principal components analysis. The characteristic band extracted were 795.815, 836.869, 885.619, 916.409, 929.239, 934.37, 957.463, 972.858, 988.253 nm; Next, matched and categorized the characteristic band of the ROI with reference spectra of the three types of garbage from the characteristic band by spectral angle mapping. The result illustrated that the classification precision of paper(A class), plastic(B class) and wood(C class) were 100%, 98% and 100% respectively and the average was 99.33%; at last, sorted the test samples by Fisher linear discrimination. The classification precision of class A, B, C were 100%, 100% and 97% respectively and the average was 99%. After a series of testing and classification by SAM and Fisher as the narrated above, the results showed that aforesaid manipulation of hyperspectral image for recyclable garbage by SAM can get more accurate results which is 99.33%. meanwhile, the research can testify that it's feasible to apply the scheme of hyperspectral imaging to assort garbage, which is significant to methodically and automatically recycle garbage in the future.
Hyperspectral imaging technology is profoundly applied into the fields of agriculture, medicine and remote sensing due to its high spectral resolution, merged image-spectrum, and fast non-destructive testing. While the method used now has the defects of long-term testing period, poor efficiency and sorting asynchrony. Spectral image can identify and classify the target garbage by establishing a recognition and classification model and analyzing reflectance spectrum information based on the facts that different materials of domestic garbage, due to their different molecular structures, will absorb different wavelengths of light and the hyperspectral image can obtain the spatial information and the reflectance spectral information from different-wavelength illumination of the target garbage. Collected recyclable garbage samples of common paper, plastic and wood materials, including plastic bottles, food packaging bags, plastic toys(jewelry) pieces, disposable chopsticks, ice cream bars, wooden furniture pieces, wooden boxes, waste textbooks, advertising paper, office paper and other items, 30 in total. And cleaned and cut them to avoid the influence of sample surface stains on the sample reflectivity. Hyperspectral imaging systems were used to acquire hyperspectral images of the sample in the near-infrared(780 ～1 000 nm) formed 18 training samples and 12 test samples. Pre-processed the collected sample image by de-noising and black-and-white correction inversion of reflectivity information. Then analyzed the region of interest of training samples by principal components analysis. The characteristic band extracted were 795.815, 836.869, 885.619, 916.409, 929.239, 934.37, 957.463, 972.858, 988.253 nm; Next, matched and categorized the characteristic band of the ROI with reference spectra of the three types of garbage from the characteristic band by spectral angle mapping. The result illustrated that the classification precision of paper(A class), plastic(B class) and wood(C class) were 100%, 98% and 100% respectively and the average was 99.33%; at last, sorted the test samples by Fisher linear discrimination. The classification precision of class A, B, C were 100%, 100% and 97% respectively and the average was 99%. After a series of testing and classification by SAM and Fisher as the narrated above, the results showed that aforesaid manipulation of hyperspectral image for recyclable garbage by SAM can get more accurate results which is 99.33%. meanwhile, the research can testify that it's feasible to apply the scheme of hyperspectral imaging to assort garbage, which is significant to methodically and automatically recycle garbage in the future.
引文
[1]Hollstein F,Wohllebe M,Arnaiz S.Identification and Sorting of Plastic Film Waster by NIR-Hyperspectral-Imaging.International Conference on Near Infrared Spectroscopy,2015.
[2]WANG Dong,LIANG Yong,MA Gang-ping,et al(王栋,梁勇,马刚平,等).Environmental Engineering(环境工程),2017,35(2):138.
[3]Lu X N,Liu F,He Z Z,et al.Science China Technological Sciences,2018,61(9):1.
[4]Yang J,Xu L,Gao M G,et al.Acta Physica Sinica,2013,62(14):140705.
[5]WEI Guo-xia,LIU Han-qiao,XU Xian,et al.(魏国侠,刘汉桥,徐仙,等)Modern Chemical Industry(现代化工),2014,34(11):46.
[6]Silveira A V M,Santana M P,Tanabe E H,et al.International Journal of Mineral Processing,2017,169(11):91.
[7]YANG Yu-qiang,WANG Kun,HUANG Huan-lin,et al(杨煜强,王坤,黄焕林,等).Acta Scientiae Circumstantiae(环境科学学报),2015,35(2):570.
[8]LIU Li-xin,LI Meng-zhu,ZHAO Zhi-gang,et al(刘立新,李梦珠,赵志刚,等).Chinese Journal of Lasers(中国激光),2018,45(2):1.
[9]GUO Xing-wang,GUAN He-qing,LIU Ying-tao,et al(郭兴旺,管和清,刘颖韬,等).Infrared and Laser Engineering(红外与激光工程),2017,46(1):51.
[10]TAN Nian,SUN Yi-dan,WANG Xue-shun,et al(谭念,孙一丹,王学顺,等).Spectroscopy and Spectral Analysis(光谱学与光谱分析),2017,37(11):3370.
[11]LIU Chang,LI Jun-wei(刘畅,李军伟).Infrared and Laser Engineering(红外与激光工程),2015,44(10):3141.
[12]WU Tong,TAN Chao(吴同,谭超).Advances in Analytical Chemistry(分析化学发展),2016,6(1):1.
[13]LIU Song-tao,CHANG Chun,MA Xin-xing,et al(刘松涛,常春,马新星,等).Laser&Infrared(激光与红外),2013,43(11):1316.
[2]WANG Dong,LIANG Yong,MA Gang-ping,et al(王栋,梁勇,马刚平,等).Environmental Engineering(环境工程),2017,35(2):138.
[3]Lu X N,Liu F,He Z Z,et al.Science China Technological Sciences,2018,61(9):1.
[4]Yang J,Xu L,Gao M G,et al.Acta Physica Sinica,2013,62(14):140705.
[5]WEI Guo-xia,LIU Han-qiao,XU Xian,et al.(魏国侠,刘汉桥,徐仙,等)Modern Chemical Industry(现代化工),2014,34(11):46.
[6]Silveira A V M,Santana M P,Tanabe E H,et al.International Journal of Mineral Processing,2017,169(11):91.
[7]YANG Yu-qiang,WANG Kun,HUANG Huan-lin,et al(杨煜强,王坤,黄焕林,等).Acta Scientiae Circumstantiae(环境科学学报),2015,35(2):570.
[8]LIU Li-xin,LI Meng-zhu,ZHAO Zhi-gang,et al(刘立新,李梦珠,赵志刚,等).Chinese Journal of Lasers(中国激光),2018,45(2):1.
[9]GUO Xing-wang,GUAN He-qing,LIU Ying-tao,et al(郭兴旺,管和清,刘颖韬,等).Infrared and Laser Engineering(红外与激光工程),2017,46(1):51.
[10]TAN Nian,SUN Yi-dan,WANG Xue-shun,et al(谭念,孙一丹,王学顺,等).Spectroscopy and Spectral Analysis(光谱学与光谱分析),2017,37(11):3370.
[11]LIU Chang,LI Jun-wei(刘畅,李军伟).Infrared and Laser Engineering(红外与激光工程),2015,44(10):3141.
[12]WU Tong,TAN Chao(吴同,谭超).Advances in Analytical Chemistry(分析化学发展),2016,6(1):1.
[13]LIU Song-tao,CHANG Chun,MA Xin-xing,et al(刘松涛,常春,马新星,等).Laser&Infrared(激光与红外),2013,43(11):1316.