Implementation based on Jetson Nano of Mung Bean Defect Classification using Image Processing and Machine Learning

Abstract

Mung beans (Vigna radiata) are a small plant species in the legume family, primarily cultivated and consumed in East Africa and South Asia. Sorting mung bean seeds is a significant challenge due to their small size. Recently, deep learning for image recognition has undergone great progress, therefore a model was developed to identify and classify mung bean seed defects with high precision and reduced processing time. In this study, 6,598 images consisting of normal and damaged mung beans were collected, and the dataset was divided into a 7:2:1 ratio for training, validation, and testing. Six object classification models were compared based on their performance during training, validation, and testing to find the best model. The results showed that MobileNetV3Large achieved the highest test accuracy of 95.75%; this was compared to MobileNetV2, EfficientNetB0, ResNet50, MobileNetV3Small, and VGG16, which achieved accuracies of 94.24%, 94.4%, 93.64%, 81.12%, and 83.64%, respectively. The weights of the best model, MobileNetV3LArge, were optimized using 16-bit floating point (FP16) quantization, combined with two-step algorithms for multiple seeds classification, and then deployed on a Jetson Nano to assess real-world performance. On the Jetson Nano, the model achieved a throughput of 3.5 Frames Per Second (FPS), maintained an acceptable accuracy of 94.85%, and a high prediction confidence once tested on real images. This research demonstrates the benefits of implementing a seed defect classification model on an embedded edge device where computational resources, time, and precision are key factors to consider.