Feature Extraction Methods on Convolution Neural Network Models for Objective Classification of Cacao Pods 

Traditional approaches of classifying cacao pods, which are usually done by hand, are often time consuming and solely relying on human judgment, which results in mixing healthy and unhealthy pods. The classification of cacao pods in the agricultural sector is crucial in determining the quality and validation of the cacao. This study presents a cacao pods classification system using Convolutional Neural Networks (CNNs) model integrated with three image feature extraction methods namely Local Binary Pattern (LBP), Edge Feature (EF), and Gabor Filter (GF). Performance of formulated models are then evaluated based on four classification metrics: Accuracy (Acc), Precision (P), Recall (R), Kappa Score (K), and F1-measure (F1). Experimental results showed that the CNN model with EF obtained the highest classification performance with Acc=0.8667, P=1.00, R=0.7872, K=0.7330, and F1=0.8485. The CNN+EF model showed a perfect predictive power in classifying unhealthy cacao pods however, it is not robust in classifying healthy cacao pods. Although this model can still be improved, its advantage is that the model is trained using images of cacao pods in the actual environment, unlike by previous scholarly works where the pods captured with uniform background. 

Keywords: cacao pods, convolutional neural network, feature extraction, local binary pattern, edge filter, Gabor filter 

Corresponding Author’s Email: oswenkylegalarce-stem@srcb.edu.ph