Robusta Coffee Bean Defect Classification using Convolutional Neural Network

Coffee consumption in the Philippines has been continuously growing but the production of green coffee beans has been decreasing since 2015, resulting in demand and supply imbalances. Two of the identified root causes are low yield and poor quality of coffee beans. Thus, this study aims to leverage image processing and Convolutional Neural Networks (CNN) to develop a deep learning model that can detect common defects (i.e., black, sour, broken, quaker, and foreign material) of Robusta coffee beans and classify them. With automated defect classification, farmers can use this data and look for trends to identify potential problems during production as coffee bean defects can be attributed to issues during planting and/or cultivation. With such data, farmers will be able to implement preventive measures early on which can eventually lead to higher overall yield and improved quality of coffee beans. Using Python, all programming, model training, and evaluation were developed through Jupyter Notebook. A modified AlexNet CNN architecture was used, and it was optimized through a series of hyperparameter tuning. Based on the results of a five-fold crossvalidation, the optimized AlexNet CNN model has an average classification accuracy of $\mathbf{92.67} \%$, F1-score of $\mathbf{92.68} \%$, and Cohen’s Kappa of $91.20 \%$.