A Proposal for Quality Classification of Cocoa Beans Using Convolutional Neural Networks

Resumo

The growing demand for technological innovations in the cacao sector and the need for standardization in the physical quality analysis of cocoa beans, given the complexity of conventional methods, motivated the development of this work. The primary goal was to build an automated cocoa bean classification tool using Convolutional Neural Networks (CNNs) applied to digital images. The proposal aims to classify cocoa beans into five specific classes, based on an image database composed of labeled samples.For performance evaluation, traditional classification metrics such as accuracy, precision, recall, F1 score, and support were adopted. Initially, three pre-trained models were used: MobileNetV2, ResNet101V2, and a network without a pre-trained model, to compare how the use of these models would influence the results. The networks were tested on two image databases: one created specifically for this study and another pre-existing, to validate the robustness of the models in different scenarios.The results were analyzed both quantitatively and visually, using graphs and confusion matrices, which facilitated the interpretation of the metrics and the performance of each network. To test the model under real-world conditions, a segmentation code based on the Otsu method was developed to identify cocoa beans through image thresholding and classify them according to the trained model.The results showed that the main network, based on MobileNetV2, achieved an accuracy of 67.14%, with an F1 score of 0.81 for the White and Defective classes, 0.60 for the Brown class, 0.49 for Partially Brown, and 0.78 for Violet. These results indicate good model performance, highlighting its potential for future mobile device applications. However, variations in the F1 scores pointed to the need for improvement in distinguishing some classes.Currently, the work is evolving with the implementation of the YOLOv12 network, an advanced pre-trained model to enhance the accuracy and speed of cocoa bean detection, benefiting from its real-time processing capability, which is essential for mobile devices and industrial environments. This study contributes to optimizing cocoa bean quality analysis, offering more precise and accessible technological solutions for the cacao industry.