Cocoa-Net: Performance Analysis on Classification of Cocoa Beans Using Structural Image Feature

The process of cocoa hybridization produces new types that have unique chemical properties that impact the manufacturing of chocolate yet are resistant to a number of plant illnesses. Image analysis is a valuable tool for visually differentiating between cocoa beans, deep learning (DL) has become the standard way for image processing. Nevertheless, these techniques necessitate a substantial quantity of data and meticulous hyperparameter adjustment. In this paper, we compare machine learning and deep learning models because it takes a lot of images to cover the wide range of agricultural products. Specifically, we extract features from images using a series of image processing techniques, and then we use both traditional machine learning methods (KNN, Decision tree, SVM, and Random Forest) and Convolutional Neural Networks (proposed Cocoa-Net and RESNET 50) to classify the cocoa beans into four categories: large, medium, small, and rejected. Since each method offers strong classification performance and has potential for use in the classification of food, they were all chosen. To test these methods, a dataset including 200 samples of fragmented images was utilised. Studies that compare various approaches are also carried out. Two optimisation techniques: Univariate Selection and Feature Importance are used to optimise the retrieved features before the machine learning deep learning models are trained. The Adam optimizer is used to optimise the proposed Cocoa-Net model. K-fold cross validation is used to assess trained models, and mean cross validation scores are then computed for performance analysis. The empirical result show that, the proposed Cocoa-Net model predicts with the highest mean accuracy score of 0.83 overall, while the Random Forest Classifier score of 0.75.