Abstract

Khao Dawk Mali 105, RD15, and RD6 are economically significant rice varieties. However, their seeds possess highly similar physical characteristics, making manual visual classification difficult. This research aims to compare the performance of two Convolutional Neural Network (CNN) architectures, MobileNetV2 and ResNet101, for the classification of these key rice varieties using a dataset of 12,000 images. The methodology involves image preprocessing by resizing images to 224x224 pixels to comply with architectural requirements. The dataset was partitioned into a 70% training set and a 30% test set to develop models using Transfer Learning. A comparison was conducted between two optimization algorithms: ADAM and SGDM. The results indicate that the ResNet101 architecture combined with the ADAM algorithm achieved the highest performance, with an average accuracy of 94.72% and an F1-score of 0.947, while MobileNetV2 achieved an accuracy of 92.00%. Specifically, the model classified Khao Dawk Mali 105 with a peak accuracy of 99.50%. Most classification errors occurred between RD15 and RD6 due to their close physical similarities. In conclusion, utilizing ResNet101 as an automated tool can reduce human error and processing time, serving as an effective prototype for quality control technology in the Thai rice industry.