Evaluation of Machine Learning Techniques for Image Based Quality Assessment of Chickpea
The quality of chickpea plays an important role to the farmers, processors, consumers, and other stakeholders. At present, the procedures for evaluating chickpea quality parameters are subjective, tedious, and destructive. The objective of this study was to develop non-destructive imaging techniques to determine chickpea quality. Eight chickpea varieties (CDC-Alma, CDC-Leader, CDC-Palmer, CDC-Frontier, CDC-Luna, CDC-Orion, CDC-Cory, CDC-Consul) were obtained from the Crop Development Centre, University of Saskatchewan and used in this research. The classification of chickpea varieties using RGB images was investigated with seven pre-trained deep convolutional neural networks (CNN) (AlexNet, GoogleNet, ResNet18, ResNet50, VGG16, VGG19, and MobileNetV2) and transfer learning. The highest overall classification accuracy of 100% was obtained for ResNet50 and MobileNetV2. The protein content in single chickpea seed for the eight varieties were predicted with hyperspectral images and chemometrics. The optimum model was developed using Partial Least Square Regression (PLSR) which yielded Correlation Coefficient of Prediction (R2p) and Root Mean Square Error of Prediction (RMSEP) values of 0.935 and 0.987. Iteratively Retaining Informative Variables (IRIV) selected wavelengths with Support Vector Machines Regression (SVMR) provided the best model with R2p and RMSEP of 0.950 and 0.857. The classification of chickpeas into hard to cook (HTC) and easy to cook (ETC- control) was carried out using hyperspectral images. Chickpeas develop HTC defect under suboptimal storage conditions resulting in extended cooking times. Support Vector Classifier (SVC) and Convolutional Neural Network-Attention (CNN-ATT) models demonstrated 100% accuracy for classifying chickpeas into HTC and ETC. IRIV selected wavelengths with SVC model yielded 100% classification accuracy. The adulteration in chickpea flour with metanil yellow was quantified with near infrared hyperspectral imaging system (NIR-HSI). Pure chickpea flour was adulterated with metanil yellow at various concentrations up to 2% (w/w). PLSR yielded a model with R2p and RMSEP values of 0.978 and 0.054 whereas One Dimensional (1D)-CNN produced a model with R2 and RMSEP of 0.992 and 0.059 for quantifying the adulterant. IRIV selected wavelengths with PLSR yielded the best model with R2 and RMSEP of 0.989 and 0.041. This research demonstrated the potential of NIR-HSI and RGB imaging systems for non-destructive and rapid determination of chickpea quality.
Saha, D., Manickavasagan, A., 2022. Chickpea varietal classification using deep convolutional neural networks with transfer learning. J. Food Process Eng. 45(3), e13975. https://doi.org/10.1111/jfpe.13975.