基于高光谱图像技术与迁移学习的水晶梨早期损伤检测

doi:10.3964/j.issn.1000-0593(2022)11-3626-05

摘要
参考文献
相关文章 (15)

全文: PDF (2205 KB)
输出: BibTeX | EndNote (RIS)

摘要：梨在储藏、包装和运输等过程中均可能发生不同程度的机械损伤，若不及时剔除损伤梨，损伤可能会逐渐严重而演变成腐烂，造成严重的经济损失。为建立一种梨早期损伤检测及损伤时间评估的快速、无损检测方法，采用高光谱图像结合迁移学习模型对损伤早期水晶梨进行识别。以无损伤、挤压损伤24 h和挤压损伤48 h的水晶梨为研究对象，应用高光谱成像系统采集样品的高光谱图像，共获取无损伤、挤压损伤24 h和挤压损伤48 h的水晶梨高光谱图像各80帧。对高光谱图像进行主成分分析，选择主成分图像4，5，6（PC4，PC5，PC6）作为检测水晶梨损伤的特征图像，将3个主成分图像拼接后进行数据扩充共得到无损伤、挤压损伤24 h和挤压损伤48 h的特征图像各160帧。按照9∶1比例划分样本训练集和测试集后，分别建立了支持向量机（SVM）、k-近邻（k-NN）和基于ResNet50网络的迁移学习损伤识别模型。SVM、k-NN和基于ResNet50网络的迁移学习模型对测试集样本总体识别准确率分别为83.33%，85.42%和93.75%，基于ResNet50网络的迁移学习模型识别效果最佳，其对测试集中无损伤、挤压损伤24 h和挤压损伤48 h的样本正确识别率分别达到100%，83%和95%。该研究结果表明，高光谱图像技术结合基于ResNet50网络的迁移学习模型可实现水晶梨早期损伤检测，并对损伤时间有较好的预测效果，且损伤时间越长，识别准确率越高。

关键词：高光谱图像；迁移学习；水晶梨；损伤检测

Abstract：Pear in storage, packaging and transportation processes may occur in different degrees of mechanical damage. If not removed in time, the damage may gradually become serious and rot, resulting in serious economic losses. In order to establish a rapid and non-destructive detection method for early bruise detection and bruise time assessment of pears, hyperspectral imaging technology combined with a transfer learning model was used to identify crystal pears with an early bruise. The hyperspectral images of crystal pears with no bruise, crush bruise at 24 h and crush bruise at 48 h were acquired by the hyperspectral imaging system, and 80 hyperspectral images of crystal pears with no bruise, crush bruise at 24 h and crush bruise at 48 h were obtained. Principal component analysis was performed on the hyperspectral images, and principal component images 4, 5 and 6 (PC4, PC5 and PC6) were selected as the feature images for detecting crystal pear bruises. After the data expansion of the stitched images with three principal components, 160 images of crystal pears with no bruise, crush bruise at 24 h and crush bruise at 48 h were obtained. The training sample set and test sample set were divided according to the ratio of 9∶1, Support Vector Machine (SVM), k-Nearest Neighbor (k-NN), and transfer learning bruise recognition models based on the ResNet50 network were established, respectively. The results showed that the overall recognition accuracy of SVM, k-NN and ResNet50 network-based transfer learning models for the test set was 83.33%, 85.42% and 93.75%, respectively. The transfer learning model based on the ResNet50 network had the best recognition results, and its correct recognition rates for the test set of crystal pears with no bruise, crush bruise at 24 h and crush bruise at 48 h reached 100%, 83% and 95%, respectively. The results of this study indicate that hyperspectral imaging technology combined with the transfer learning method based on the ResNet50 network can achieve early bruise detection of crystal pears and have a great prediction performance for bruise time, and the longer the bruise time, the higher the recognition accuracy.

Key words：Hyperspectral imaging; Transfer leaning; Crystal pear; Bruise detection

收稿日期: 2021-08-30 修订日期: 2022-04-05

中图分类号:

TP751

基金资助: 中国博士后基金面上项目(2021M690573)，中央高校基本科研业务费专项资金项目(2572020BF05)资助

通讯作者: 刘大洋 E-mail: ldy333ldy@163.com

作者简介: 王广来，1996年生，东北林业大学机电工程学院硕士研究生 e-mail: gellerwang@163.com

引用本文:

王广来，王恩凤，王聪聪，刘大洋. 基于高光谱图像技术与迁移学习的水晶梨早期损伤检测[J]. 光谱学与光谱分析, 2022, 42(11): 3626-3630.
WANG Guang-lai, WANG En-feng, WANG Cong-cong, LIU Da-yang. Early Bruise Detection of Crystal Pear Based on Hyperspectral Imaging Technology and Transfer Learning. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2022, 42(11): 3626-3630.

链接本文:

https://www.gpxygpfx.com/CN/10.3964/j.issn.1000-0593(2022)11-3626-05 或 https://www.gpxygpfx.com/CN/Y2022/V42/I11/3626

[1] GUO Wen-chuan, WANG Ming-hai, YUE Rong(郭文川, 王铭海, 岳绒). Transactions of the Chinese Society for Agricultural Machinery(农业机械学报), 2013, 44(2): 142.
[2] Zhang M, Li C, Takeda F, et al. Transactions of the ASABE, 2017, 60(5): 1489.
[3] Pan X Y, Sun L J, Li Y S, et al. Journal of the Science of Food and Agriculture, 2019, 99(4): 1709.
[4] SUN Shi-peng, PENG Jun, LI Rui, et al(孙世鹏, 彭俊, 李瑞, 等). Food Science(食品科学), 2017, 38(2): 301.
[5] CHEN Xin-xin, GUO Chen-tong, ZHANG Chu, et al(陈欣欣, 郭辰彤, 张初, 等). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2017, 37(1): 150.
[6] Jiang H, Zhang C, He Y, et al. Applied Sciences-Basel, 2016, 6(12): 450.
[7] TIAN You-wen, WU Wei, LU Shi-qian, et al(田有文, 吴伟, 卢时铅, 等). Food Science(食品科学), 2021, 42(19): 260.
[8] Zhang M Y, Jiang Y, Li C Y, et al. Biosystems Engineering, 2020, 192: 159.
[9] Wang Z D, Hu M H, Zhai G T. Sensors, 2018, 18(4): 1126.
[10] Azizah L M, Umayah S F, Riyadi S, et al. Deep Learning Implementation Using Convolutional Nerual Network in Mangosteen Surface Defect Detection. 2017 7th IEEE International Conference on Control System, Computing and Engineering (ICCSCE), 2017. 242. doi: 10.1109/ICCSCE.2017.8284412.
[11] Bai Yuhao, Xiong Yingjun, Huang Jichao, et al. Postharvest Biology and Technology, 2019, 156: 110943.
[12] WEI Chen-jie, WANG Ji-fen, ZENG Xiao-hu(卫辰洁, 王继芬, 曾啸虎). Journal of Instrumental Analysis(分析测试学报), 2021, 40(7): 1043.
[13] YU Xiao-na, HUANG Liang, CHEN Peng-di(余晓娜, 黄亮, 陈朋弟). Journal of Chongqing University(重庆大学学报), 2022, https://kns.cnki.net/kcms/detail/50.1044.N.20210615.1554.005.html.
[14] He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 770. doi: 10.1109/CVPR.2016.90.
[15] WANG Chun-shan, ZHOU Ji, WU Hua-rui, et al(王春山, 周冀, 吴华瑞, 等). Transactions of the Chinese Society of Agricultural Engineering(农业工程学报), 2020, 36(20): 209.
[16] LIN Si-han, LI Jing, XUE Long, et al(林思寒, 黎静, 薛龙, 等). Acta Agriculturae Universitatis Jiangxiensis(江西农业大学学报), 2018, 40(4): 835.