Object Detection Model for Indonesian Food Using Machine Learning

Some of Indonesia’s biggest health problems have a direct correlation to Obesity. Because of this, it is important to raise awareness about how to control eating habits, one of which is the use of Dietary Self-Monitoring (DSM) Applications. The prediction of food classes is well documented and has been widely used in various DSM applications. However, the applications do not have sufficient data and trained models to be used to detect Indonesian foods. This study presents a Machine Learning based approach to build an object detection model that would be able to accurately detect Indonesian Foods. The system was trained with three different methods of object detection, which consists of Faster Region-Based Convolutional Neural Network (R-CNN), Single-Shot Detector (SSD), and You Only Look Once v5 (YOLOv5) with thousands of Images that were manually collected. After testing, all three models achieved good performances, being able to reach a 0.7 Mean Average Precision (mAP).5:.95 score across the board. However, one model was clearly above the others, which was the YOLOv5. After further testing of the YOLOv5 model, it was then modified to further improve results. The Modified YOLOv5 model then managed to achieve an impressive 0.811 mAP.5:.95 score. In conclusion, the Modified YOLOv5 model is the best fit for our dataset based on the performance.