Performance Analysis of EfficientNet for Rice Grain Quality Control-An Evaluation against YOLOv7 and YOLOv8
Gilbert G. Yadao, Omar Mukhtar Y. Julkipli, Cyrel O. Manlises
Abstract
This study presents a comprehensive performance analysis of EfficientNet, focusing on its application in rice grain quality control and comparing it with two other deep learning algorithms, YOLOv7 and YOLOv8. The objective is to evaluate the effectiveness of EfficientNet in classifying rice grains into four categories: Healthy, Broken, Damaged/Discolored, and Chalky. The research utilized an RPI 4B with a 12.3-megapixel Raspberry Pi HQ camera module with a lens for data acquisition. Training data consisted of jpg/jpeg images of rice grains collected from various sources, including healthy kernels from Thailand, broken rice from Vietnam, and damaged/discolored and chalky grains from Region IV – Mindoro. YOLOv7 attained an 84.80% accuracy, YOLOv8 achieved 87.37%, while the EfficientNet model excelled with the highest accuracy at 90.39%. This highlights EfficientNet's superior performance in rice grain classification.