Convolutional Neural Network and Industrial Robot Arm applied to an automatic coffee bean selection system

Abstract

The objective of this paper is to design and implement an automatic coffee bean selection system, based on the integration of a Scara Epson robot arm with a Convolutional Neural Network - based classifier. The implemented system extracts the elements that were identified as coffee beans with shape and color alterations. The hardware architecture consists of: Epson Scara LS10 robot arm, RC-90B controller, 4-megapixel webcam, an extraction end effector, and white illumination. The software architecture consists of: image acquisition, segmentation and preprocessing algorithms, training and classification algorithms with Convolutional Neural Networks (224/2 input/output layers), and robot arm motion control algorithms. For the performance evaluation of the automatic classification algorithms, 18 tests were performed considering 3 different cases of separation between grains, greater than 5 mm, 3 to 4 mm, and less than 2 mm. As a result, an effectiveness percentage of 100% was obtained for the first and second range of separation, and a percentage of 61.5% for the third range, due to the overlapping between coffee beans.