Real-Time Detection of Rice Quantity in Institutional Food Service Using the DeepLabV3Plus Deep Learning Algorithm

This study aimed to identify the most suitable algorithm for accurate and efficient real-time rice quantity detection on white serving trays to enhance personalized nutrition management and improve dietary quality and health. A real-time rice quantity detection system was developed using the DeepLabV3Plus network with various backbone architectures, including DenseNet121, DenseNet201, VGG16, VGG19, MobileNetV2, and MobileNetV3 Large. DeepLabV3Plus employed Atrous Convolutional Neural Networks (ACNNs) and Atrous Spatial Pyramid Pooling (ASPP). A total of 350 images of rice servings were used, with 70% for training and 30% for validation. Performance evaluation of the six backbone architectures revealed that DenseNet121 and DenseNet201 achieved the highest Mean Intersection over Union (mIOU) of 92.92%, while MobileNetV2 had the lowest at 87.70%. VGG16 demonstrated the highest computational efficiency. DenseNet201 was effective in detecting rice quantities up to 140 g, but accurate detection of larger quantities required additional height image data. All architectures achieved high top-1 validation accuracies (96.0%-96.4%), though training times varied significantly. This study emphasizes the importance of diverse training datasets and continuous model refinement to improve segmentation accuracy and efficiency. Integrating image data from top and side views is essential for detecting larger quantities. Enhanced nutrition management using deep learning in institutional food services could lead to increased operational efficiency, cost savings, and data-driven decision-making for inventory and menu planning.