高光谱成像的草莓缺陷检测及可视化  

Strawberry Defect Detection and Visualization Via Hyperspectral Imaging

作　　者：赵路路 周松斌 刘忆森 庞锟锟 殷泽轩 陈红 ZHAO Lu-lu;ZHOU Song-bin;LIU Yi-sen;PANG Kun-kun;YIN Ze-xuan;CHEN Hong(Guangdong Institute of Intelligent Manufacturing,Guangzhou 510000,China;Guangdong Key Laboratory of Modern Control Technology,Guangzhou 510000,China)

机构地区：[1]广东省科学院智能制造研究所,广东广州510000 [2]广东省现代控制技术重点实验室,广东广州510000

出　　处：《光谱学与光谱分析》2025年第5期1310-1318,共9页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金面上项目(62275056);广东省自然科学基金面上项目(2024A1515011628);广东省科学院发展专项资金项目(2022GDASZH2022010108)资助。

摘　　要：草莓在采摘、运输、贮藏、包装和销售过程中容易造成不同程度损伤与缺陷,如淤伤、冻伤,真菌感染等,给果农和销售商带来较大的经济损失。高光谱技术是光谱传感与视觉技术的结合,可实现水果各类品质缺陷的无损检测。然而,目前高光谱水果检测的建模方法仍存在两个方面的问题:首先,输入信息主要采用平均光谱为主,对高光谱的图像信息利用不足;其次,目前卷积网络已成为高光谱信息处理的发展趋势,但卷积网络存在感知域较小,难以获得谱段或图像信息的长程关系。为解决上述问题,实现多种草莓缺陷的准确检测与识别,提出空间光谱变换网络(SSTN)对四类(健康、瘀伤、冻伤、感染)草莓的近红外高光谱数据(900~1700 nm)进行分类。SSTN以Vision Transformer(ViT)网络为主体,将高光谱数据块进行位置编码作为输入信息,从而实现“图谱联合”建模,其内部的多头注意力机制还可捕获长距离谱段/图像关系。实验方面,以128个健康、128个淤伤、128个冻伤、118个感染,共计502个草莓作为样本,按照1∶1的比例随机划分训练集和测试集,进行分类建模实验。结果显示,SSTN模型的分类准确率最高,达到99.20%,相比于一维卷积神经网络(1D-CNN)、二维卷积神经网络(2D-CNN)和注意力卷积网络(CBAM-CNN),精度分别提升了3.8%、3.3%及1.5%。为了能够进一步可视化各类草莓缺陷的具体位置,将训练好的2D-CNN、CBAM-CNN和SSTN模型分别与Score-CAM结合进行可视化。缺陷可视化结果显示,CBAM-CNN模型中的卷积注意力机制能够提升缺陷定位的准确性,而具有多头注意力机制的SSTN模型结合Score-CAM获得最佳的可视化效果,能够准确的显示出缺陷的位置和缺陷形状轮廓。该研究为建立一种快速、无损、自动化的草莓缺陷检测方法提供参考。Strawberries can be easily damaged during harvesting,transportation,storage,packaging,and sales.The damages and defects encountered include bruising,frost damage,and fungal infections,which can cause great economic losses to fruit farmers and sellers.Hyperspectral technology combines spectral sensing and machine vision to non-destructively detect various quality defects in fruits.However,there are currently two problems in modeling hyperspectral fruit detection:First,the input information is mainly based on average spectra,and the hyperspectral image information is not adequately utilized.Secondly,convolutional networks(CNN)have become the main focus of development in hyperspectral data processing.Still,CNNs have a relatively narrow domain of perception,and it is difficult to obtain long-term correlations for spectral segments or image information.To solve the above problems and accurately detect and recognize various strawberry defects,a spatial-spectral transformation network(SSTN)was proposed to classify the near-infrared hyperspectral data(900~1700 nm)of four categories of strawberries(healthy,bruised,frost damaged,and infected).The SSTN uses the Vision Transformer(ViT)network as the main body and hyperspectral data patches with encoded position information are used as inputs to achieve“spectra-spatial”modeling.The model's internal multi-head attention mechanism can also capture long-distance spectral/spatial correlations.In the experiment,502 strawberries were sampled,including 128 healthy,128 bruised,128 frost damaged,and 118 infected strawberries.The training and test sets were randomly divided according to a1∶1 ratio.Half of the data was used to train the model to classify defects,and the other half was used to test the model's performance.The results show that SSTN achieved a maximum classification accuracy of 99.20%.Compared with one-dimensional convolutional neural network(1D-CNN),two-dimensional convolutional neural network(2D-CNN),and convolutional network with attention mechanism(CBAM-CNN),our

关 键 词：草莓缺陷 高光谱 Score-CAM 可视化 

分 类 号：O657.3[理学—分析化学]