Evaluation of YOLO Efficiency in Automatic Orange Detection in Multi-Exposure Images

Abstract

Brazil is the largest producer of oranges in the world and the automatic detection of fruits has been a challenging task in the context of remote sensing, due to variations in fruit appearance, changes in lighting and occlusions of foliage and neighboring fruits. In this sense, this paper focus on the detection of oranges in multispectral images, with different spectral bands and exposures, using a convolutional neural network (CNN) known as YOU ONLY LOOK ONCE (YOLO). The results indicate that, after 300 epochs, the model demonstrated an accuracy of 81.5% and an approximate recovery rate of 85%. Shutter speeds 1/640s and 1/250s are not suitable for detection due to low light and overexposure, respectively. Intermediate values may be more suitable for identifying a larger number of fruits.