Image-based Quality Identification of Black Soybean (Glycine soja) Using Convolutional Neural Network

Abstract

The problem faced in identifying the quality of black soybeans is that the quality of the assessment is inconsistent and it takes a relatively long time. This study aims to determine the best convolutional neural network architecture by comparing the performance of Custom CNN, MobileNetV2, and ResNet-34 architectures in identifying the quality level (grade) of black soybeans. The quality of black soybean is split into 4 different classes based on physical characteristics (split, damaged, other colors, wrinkles, dirt) and moisture content test. The number of images used is 1300 images, with the ratio of training data, validation data, and testing data are 50:25:25, 60:25:15, and 70:20:10. The best model for identifying the quality based on the physical characteristics is the MobileNetV2 architecture with a ratio of 50:25:25 which produces an accuracy of 90.18%. Morover, the best model for identifying the quality based on the moisture content is the ResNet-34 architecture with a ratio of 70:20:10, which produces an accuracy of 78.12%. The best overall accuracy in identifying the quality based on both physical characteristics and moisture content is the ResNet-34 architecture, with a ratio of 70:20:10, with an average accuracy of testing data of 79.21%.