薄层色谱结合深度学习对茜草染料的智能识别OA

Rubia cordifolia,as the most important red natural plant dye,is widely used.There are numerous species of Rubia plants,and significant differences in their chemical compositions due to variations in habitats.However,existing detection methods struggle to achieve rapid and accurate analysis and classification.The purpose of this study is to establish a technical system combining thin layer chromatography(TLC)with deep learning,solve the problem of rapid identification and classification of Rubia cordifolia from different habitats,provide technical support for its quality evaluation and origin traceability,and offer reference for the intelligent detection of natural dyes. The study used Rubia cordifolia samples from eight habitats,including Xinjiang,India,and Tibet,as research objects.Firstly,the TLC experimental conditions were optimized.Extracts of Rubia plants from the eight regions were subjected to extensive repeated TLC under optimized conditions,and images were collected.Subsequently,the processed TLC strip images were used to train eight deep learning models,such as ResNet,DenseNet,and ViT(Vision Transformer),with performance evaluated.Finally,Grad-CAM(gradient-weighted class activation mapping)was used to visually analyze the model's decision-making basis.This study integrates the advantages of TLC(simple operation and low cost)with the powerful feature mining capability of deep learning,optimizes the TLC conditions for Rubia cordifolia,and evaluates the classification performance of eight network models for Rubia cordifolia from different habitats,thereby realizing the variety identification and origin classification of Rubia cordifolia.Through investigations of four commercial thin-layer plate brands,developing solvent systems,and sample volumes,it was found that the Huanghai silica gel G plate was most suitable for this experiment(Fig.1).A sample volume of 5 μL(Fig.3)yielded chromatograms with clear spots and good resolution when developed with petroleum ether(60-90℃)-acetone(5:2,v/v)as the developing solvent and visualized under a 365 nm ultraviolet(UV)lamp(Fig.2).The samples showed fluorescent spots consistent with reference substances(alizarin,purpurin,rubicin,etc.)(Fig.4).By comparing the classification performance indicators of eight deep learning models(ResNet,DenseNet,ViT,Inception V3,etc.),the CBAM-DenseNet201 model achieved the best classification effect for Rubia cordifolia,with an accuracy(ACC)of 98.21%,area under the ROC curve(AUC)of 98.57%,and F1-score of 92.90%.The Inception V3 model exhibited the poorest classification performance,with an ACC of 89.64%and an F1-score of 57.81%.ViT,ResNet18,ResNet34,and their CBAM(convolutional block attention module)variants also demonstrated good classification accuracy,with ACC exceeding 93%.Additionally,the performance indicators of the CBAM variants of ResNet18,ResNet34,and DenseNet201 were superior to their original models(Tab.2).Grad-CAM visualization showed that the three attention-enhanced models(CBAM-ResNet18,CBAM-ResNet34,CBAM-DenseNet201)mainly focused on the fluorescent spot regions in the TLC images,indicating that the models recognized and classified Rubia cordifolia based on spot characteristics rather than irrelevant backgrounds.Different models focused on different regions in the TLC images of Rubia cordifolia from different habitats:the CBAM-DenseNet201 model showed the highest correspondence between its focus regions and actual spots,while ResNet series models occasionally focused on background regions without spots.Furthermore,the three models focused on similar spot regions in the TLC images of Rubia cordifolia from the same habitat,indicating that different network structures can capture similar key features(Fig.5).The classification basis of the models is consistent with the chemical significance of TLC:the Rf(retention factor)value of spots reflects component polarity,and spot intensity reflects component content.This confirms that the models'focus on spot regions enables deep learning to capture the chemical characteristics of TLC,achieving rapid and effective mapping from images to component differences. In this study,deep learning technology was applied to the analysis of TLC fingerprints of Rubia cordifolia.However,the experiment only selected samples from eight habitats,covering a limited range of varieties and geographical distributions,which cannot fully reflect the diversity of global Rubia species.In future studies,we will enrich Rubia cordifolia samples from additional habitats to improve the model's adaptability to complex real-world scenarios.We will also attempt to extend the TLC-combined deep learning technology to the analysis of other natural dyes(e.g.,flavonoids,alkaloids,indoles),so as to provide basic research for the standardization and intelligent analysis of natural dyes.