Hybrid methods for detection of starch in adulterated turmeric from colour images

In this investigation, an image dataset featuring starch-adulterated turmeric was meticulously created. Various concentrations of starch, ranging from 0 to 100%, were blended with turmeric powder and documented using a digital camera. The study proposed a hybrid methodology, amalgamating a modified pre-trained Transfer Learning (TL) deep convolutional neural network (DCNN) known as DenseNet201 with a feature selection technique and machine learning (ML) classifiers. The basic DenseNet201 model was modified by adding a few layers. This modified DenseNet201 functioned as a feature extractor, and optimal attributes were chosen using the ExtraTreeClassifier (ETC) method from these extracted features. Diverse ML models were employed to forecast starch concentration levels based on the ETC-selected features. The investigation encompassed the development of both classification and regression models for starch prediction. The Grad-CAM visualization method was applied to visualize class activation maps.

The proposed system (modified Densenet201-ETC), integrating Logistic Regression (LR) and Decision Tree Classifier (DTC), demonstrated remarkable results: 100% accuracy, an Area Under The Curve-Receiver Operating Characteristics (AUC-ROC) score of 1.0, and a 98.3% average accuracy over the 5-fold cross-validation for the classification method. The regression method, implemented with K-Nearest Neighbors Regressor (KNNR), achieved a coefficient of determination (R2) of 0.97, root mean square error (RMSE) of 0.65, leave-one-out cross-validation (LOOCV)-R2 of 0.90, and an average LOOCV-RMSE of 1.09. Moreover, the performance of five existing TL models, DenseNet201, Xception, ResNet50, InceptionV3, and VGG16, compared with the proposed method.