Vegetable Mass Estimation based on Monocular Camera using Convolutional Neural Network

Abstract

Vegetable mass estimation from monocular RGB camera images is proposed. Vegetables are fragmented and placed on a conveyor belt of food processing machine and the monocular camera placed over the belt take pictures of vegetables on the belt. The proposed system does not employ any scale, load cell, and other mass scaling equipment. We apply pre-trained convolutional neural networks to estimate the mass of vegetables. Transfer learning including various levels of fine-tuning is also applied. For pre-trained network, we use Xception, VGG16, ResNet50, and Inception_v3, which are pre-trained using ImageNet. The result shows that the best estimation accuracy is achieved by VGG16, whose MAPE (mean average percentage error) is 11.1%. Additionally, we fine-tune VGG16 and the accuracy reduces to 7.9% for MAPE. From this result, the performance of CNN model can improve by fine-tuning. The proposed system can be applied to low-cost, high-speed, and efficient measurement of foods replaced to load cells.