Benchmarking A New Dataset for Coffee Bean Defects Classification Based on SNI 01-2907-2008

To develop a device which can determine automatically the quality of coffee beans, the module to identify and classify coffee bean defects based on digital images processing is needed. For this reason, a new dataset of coffee bean defect images is built in this research. 3 classes were defined and obtained from the image sample of coffee beans, namely good coffee beans, defective coffee beans, and non-valid coffee beans. Moreover, specifically for the class of defective coffee beans, the new dataset also provides ground truth image samples for 17 classes of coffee bean with defects which was defined by SNI 01-2907-2008. Two Convolutional Neural Network (CNN) architecture models, InceptionResnetV2 and MobileNet, were tested to do an initial benchmarking process to classify coffee beans and their types of defects. Both architectures show very promising results to classify coffee bean images in 3-classes. With 92.52% of accuracy, MobileNet outperforms slightly InceptionResnetV2. Both architectures are fairly easy to classify the 3-classes of coffee bean images. But the experimental results indicate that the task of classifying more specifically the 17 classes of different types of defects in coffee beans is still quite difficult. Although the accuracy of InceptionResnetV2 of 53.35% seems much higher than the accuracy of MobileNet, the two architectures still seem to be struggling to identify each type of coffee bean defect. This benchmarking results will serve as the initial basis that helps to analyze the level of complexity of problems in the coffee bean new dataset, as well as the basis for further investigations and research processes to improve the performance of the future coffee bean classifier machine.