A coffee bean classifier system by roast quality using convolutional neural networks and computer vision implemented in an NVIDIA Jetson Nano

This article proposes the development and implementation of an intelligent system for the automatic classification of coffee beans by roast level or quality, which complements a previous project on the development of a coffee bean detector. Likewise, this work, in subsequent articles, will be part of a more complex system that will carry out the selection of beans by means of electro-pneumatic actuators. According to different study sources, currently, the coffee agro-industrial sector employs labour to discern the quality of the beans that manage to go through the selection process to the final packaged product. The expenses of the coffee growing companies are considerably affected by the investment in personnel for the selection of bean, in addition, the results of this are not always exact and uniform, since they are influenced by the subjectivity in the vision, and in the criteria or judgment, of each operator. For this reason, this work presents as a solution a classification system for coffee beans according to their level of roasting, which is highly linked to their final quality. This system was developed in Python and implemented on an NVIDIA Jetson Nano development board, where computer vision libraries such as OpenCV and artificial intelligence libraries such as Pytorch were used, the latter to design a convolutional neural network (CNN) and train it with our own dataset, obtained from real samples of coffee beans differentiated into 3 degrees of roasting (under-roasting, optimum roasting and over-roasting).