Multi-label Coffee Bean Classification Using Deep Learning

Abstract

Coffee is a critical agricultural product and a significant economic commodity, particularly in Ethiopia, where it serves as a major export. Traditional methods of coffee bean assessment are labor-intensive and prone to human error, necessitating the development of automated, accurate solutions. This study addresses this problem by applying deep learning techniques, specifically Convolutional Neural Networks (CNNs), to classify coffee beans based on their quality and geographical origin. The primary objective of this research is to design and develop a robust CNN model capable of accurately classifying coffee beans using image data. To achieve this, we collected an extensive dataset of 3,965 coffee beans from the Ethiopian regions of Arsi, Yirgacheffe, Guji, and Sidama, photographing each bean from the front and back to obtain a total of 6,373 images. To enhance the dataset's balance, an additional 2,417 images were generated through augmentation, bringing the final dataset to 8,790 images. The methodology involved preprocessing these images, followed by training and evaluating multiple CNN architectures. Techniques employed include the use of both grayscale and RGB image data, Dropout layers to prevent over-fitting, and the Adam optimizer for efficient training. Our results indicate that the CNN model trained on RGB images achieved a peak accuracy of 99.66% in a specific experiment, while the average accuracy across all experiments was 99.06%. This highlights the significance of color information in enhancing feature extraction and model learning. The model's high accuracy and efficiency demonstrate its potential to automate and improve coffee quality assessment, offering a valuable tool for the coffee industry. By leveraging deep learning techniques, this study provides a scalable and precise solution for the classification of coffee beans, potentially transforming quality control processes and setting new standards in the coffee industry. The study's findings underscore the superiority of CNN-based models in handling complex image data, particularly when color features play a crucial role in classification tasks