Coffee Bean Quality Classification Using the Variational Autoencoder and Support Vector Machine Algorithms

Abstract

The conventional process for sorting coffee bean quality on a large scale in the downstream coffee industry is still prone to human error, labour-intensive, and time-consuming. This research utilises autoencoder and support vector machine (SVM) techniques to develop a coffee bean quality classification model with four categories: ripe, unripe, non-coffee, and damaged beans. Additionally, this study aims to evaluate the model's performance using F1 score metrics, confusion matrix analysis, and k-fold cross-validation. A total of 1,000 image samples were used and evenly distributed into four classes. This was then further divided into a training subset of 70%, a validation subset of 20%, and a testing subset of 10%. The proposed model achieved a classification accuracy of 96%, demonstrating a high level of reliability. Additionally, the weighted average F1 score of 0.961 indicates a strong balance between precision and recall values across all classes. VAE + SVM outperformed the baseline CNN in classifying coffee bean quality, increasing accuracy by 3% from 93% to 96%. These results confirm the model's ability to effectively predict coffee bean quality and highlight its potential as a robust solution for automating the sorting process in real-world coffee.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint