Classification of Coffee Beans Defect Using Mask Region-based Convolutional Neural Network

Abstract

The recognition and calculation of coffee bean defects has become one of the references in determining the quality of coffee beans. Indonesia already has a standard set forth in Indonesian National Standard (SNI) to determine the quality of coffee beans based on the special quality requirements of beans by calculating the value of defects. In the standard there are several types of seed defects defined such as black seeds, hollow seeds, cracked seeds and others. The system for recognizing and calculating the value of coffee bean defects is still done conventionally so that it requires a very high ability and skill from the examiner. The purpose of this study is to design a model that is able to recognize coffee bean defects using one of the models in the Deep Learning approach, namely Convolutional Neural Network (CNN). This study will use an improvised workflow from CNN, namely Mask R-CNN (Region-based Convolutional Neural Network) which utilizes the workflow of Faster R-CNN with the addition of a masking feature on the detected object. This study uses a dataset of 480 images with black, broken and hole classes with 360 images for training and 120 images for validation. The test is carried out with 2 forms of images, namely images with individual objects with 30 images and images with plural objects with 20 images. The accuracy obtained is 93.3% for testing individual objects and 75% for testing plural objects.