The Effect of Types and Amount of Training Labels on the Accuracy of CNN for Yuzu Fruit Recognition

Abstract

In order to realize automatic counting of fruit using deep learning, construction of training label is required. However, the effect of labeling on object recognition models is not well understood, and there is also uncertainty regarding the amount of training data required for these models. Therefore, this study aims to investigate the effect of different labeling approaches on classification performance by assigning different labels to fruits with different ripeness levels and partially hidden fruits. Classification models are constructed using different datasets, and metrics such as F1 score, mAP50 are compared. Regarding the consideration of GPU memory usage and training computation time, model size “m” is most suitable. Additionally, practical performance assurance can be achieved with a dataset of around 100 to 150 images. The study reveals that, for data sizes less than 100, model size “s” performs well, while for data sizes exceeding 100, model sizes “m,” “l,” and “x” exhibit higher performance. Furthermore, in practical scenarios, a dataset size of around 100 to 150 images is sufficient, and increasing the data size further leads to performance improvement for model sizes other than “x.”