Classification of fresh cocoa beans with pulp based on computer vision

Precision farming is one of the most trending topics nowadays and computer vision techniques are increasingly gaining momentum on this subject. On the other hand, in the cocoa industry, particularly in small farms, farmers still perform the classification of fresh cocoa beans with pulp in a traditional way, i.e. through their senses. In this work, we explain a new approach for cocoa beans with pulp classification, in order to aid in the process of removing cocoa beans pulp to efficiently estimate the quality of these beans. Our approach used morphological operations, k-means clustering, a bag of visual words as a feature extractor, and finally a support vector machine classifier. We achieved an AUC of 97.757% and an accuracy of 97.57% with a low false-positive rate of 2.46%, which demonstrates the viability of using computer vision for this task. We used a real-world dataset of 247 fresh cocoa beans images, that we collected and labeled with experienced cocoa farmers.