Fruit detection in viticulture with deep neural networks

We investigate deep neural networks applied to fruit detection in viticulture. We also developed the Embrapa WGISD dataset, composed of images collected in April 2017/2018 at the Guaspari Winery. Annotated manually, the dataset has 5 different varieties of grape: Syrah, Chardonnay, Cabernet Franc, Cabernet Sauvignon, and Sauvignon Blanc, totaling 4419 samples of grape bunches. We trained YOLOv2 and YOLOv3 to detect and locate the bunches in the images. YOLOv2 achieved up to 88% accuracy and YOLOv3 92% accuracy. Qualitative tests demonstrated that the YOLOv2 network generalizes better for the dataset used, and the YOLOv3 network provides a better-adjusted location.