Coffee Bean High Accuracy Classification with eXplainable Artificial Intelligence

With the development of exploration and trade, the coffee industry blossomed, and coffee gradually became an essential social drink for its unique taste and captivating aroma. Among various industrial coffee-producing processes, coffee bean classification is of great importance, which involves classifying beans into different qualities as multiple kinds of beans contain distinct characteristics, such as fermentation, sucrose, and sour degrees, influencing coffee beverage quality control, price setting, and even consumers' health. To the best of our knowledge, a prevailing approach for classification is machine learning based on neural networks. Unfortunately, even though high accuracy could be accomplished, it is hard to convey convincing results to researchers or industry personnel due to the 'black-box' property of the neural network, which indicates that we have little information about the inner mechanism in the neural network except for its inputs and outputs, jeopardizing the transparency and reliability of the neural network. Excellent interpretability is meaningful for retailers and consumers to know why the outputs are reliable. To begin with, in this paper, we will stress the importance of the high-accuracy classified coffee bean. Secondly, we will introduce the deep neural network technique to improve its high-accuracy classification. Last, we will demonstrate the XAI technique (GradCAM, etc.) to enhance explainability and interpretability. We improved the interpretability of the heatmap results produced by GradCAM and GradCAM++, and we also found out that they are presented differently. As XAI approaches, the versatile GradCAM provides high-resolution heatmaps to assist in understanding the regions of pictures that contribute much to the classification results, and the GradCAM++ produces much more precise results on accurately locating the contributing regions, including a better presentation of the unique distinct features of each class.