基于图像光谱超分辨率的苹果糖度检测

Prediction of Soluble Solid Content in Apple Using Image Spectral Super-Resolution

苹果风味独特,清脆可口,深受全世界消费者的广泛喜爱。糖度是衡量苹果品质的关键指标。高光谱成像(HSI)由于含有丰富的图谱信息在糖度无损检测中有着广泛的应用前景,然而仍面临仪器笨重昂贵、操作耗时等问题。光谱超分辨率(SSR)可通过建立映射关系从低光谱维度RGB图像获得对应高光谱维度HSI图像,在HSI图像的高效获取上有着极大的优势。因而,将探索苹果RGB图像的SSR,并基于SSR数据进行糖度预测。首先,选取大小均匀的苹果作为研究对象,利用黑色哑光胶纸对感兴趣区域(ROI)进行标定。采集苹果RGB图像和HSI图像后,利用全局阈值法确定ROI并经过图像分割得到220个RGB-HSI图像对。然后,使用密集连接网络、多尺度层级回归网络和Transformer网络实现苹果RGB图像的SSR。最后,提取SSR后图像的反射率光谱,采用全光谱和竞争性自适应重加权选择后的有效波长光谱结合偏最小二乘回归(PLSR)、随机森林(RF)和极限学习机(ELM)构建糖度预测模型。结果表明,基于Transformer网络SSR结果最好。在SSR预测集中,平均相对绝对值(MRAESP)为0.1359,均方根误差(RMSESP)为0.0262;SSR后方法的反射率光谱与真实光谱一致性最好。在糖度预测的过程中,全光谱下ELM模型预测效果最好,预测集决定系数(RP2)和均方根误差(RMSEP)为0.9255和0.003,PLSR次之,RF最差。经过有效波长光谱提取后,预测结果有所提升,其中ELM模型预测结果最好,RP2为0.9609,RMSEP为0.0022,PLSR次之,RF最差。总之,基于Transformer图像SSR完成了苹果糖度的准确检测,提供了低成本高效率HSI图像的获取方式,实现了快速便捷的新型糖度检测,扩展了图像在水果品质分析中的应用场景,为促进智慧农业和食品领域的发展提供了理论依据。

Apples have a unique flavor,crisp and delicious,and are widely loved by consumers worldwide.Soluble solid content(SSC)is an important internal quality indicator of apples.Hyperspectral imaging(HSI)has been widely used as a nondestructive tool to predict SSC in apples because it can simultaneously acquire spatial and spectral information.However,the widespread application of HSI is hindered due to expensive equipment and time-consuming operations.Spectral super-resolution(SSR)is an efficient way to acquire HSI images by establishing a mapping relationship from low spectral resolution images to corresponding high spectral resolution images.Hence,this study aims to adopt SSR to obtain HSI images from apples RGB images and use the hyperspectral data to predict the SSC of apples.Firstly,the apples of uniform size are selected as samples.Each apple is marked using the black grid matte paper to label the region of interest(ROI),and RGB and HSI images of apples are measured.Then,the global thresholding method generates 220 ROI image pairs of RGB and HSI.Secondly,a dense connection network,a multi-scale hierarchical regression network,and a Transformer network are used to achieve SSR of Apple RGB images to gain HSI images.Finally,the reflectance spectra of HSI images were extracted,and a competitive adaptive reweighted sampling algorithm was applied to obtain the spectra of effective wavelengths(EWs).Partial least squares regression(PLSR),random forest(RF),and extreme learning machine(ELM)are used to predict the SSC of apples by using the full spectra and spectra of EWs.The results show that the Transformer network achieves the best SSR with the mean relative absolute error(MRAE_(SP))of 0.1359and the root mean square error(RMSE_(SP))of 0.0262in the SSR prediction set,and the spectra obtained after SSR are most consistent with the ground truth.As for the full spectra,ELM provides the best prediction performance for SSC analysis with the coefficient of determination(R_P~2)of 0.9255and root mean square error(RMSE_P)of0.003in the prediction set.The prediction results of PLSR were relatively poor,and RF performed the worst.When the spectra of EWs are used,tELM obtains the optimal performance R_P~2=0.9609and RMSE_P=0.0022.In contrast,PLSR obtains a slightly poor result and the worst result of estimating SSC is acquired by RF.In conclusion,based on the Transformer image SSR,this article has accomplished the accurate detection of sugar content in apples,offering a low-cost and efficient method for obtaining HSI images.It has realized a rapid and convenient new sugar content detection method,expanding the imaging application scenarios in fruit quality analysis.This provides a theoretical basis for promoting the development of smart agriculture and the food industry.