Automatic Differentiation of Damaged and Unharmed Grapes Using RGB Images and Convolutional Neural Networks

摘要

Knowledge about the damage of grapevine berries in the vineyard is important for breeders and farmers. Damage to berries can be caused for example by mechanical machines during vineyard management, various diseases, parasites or abiotic stress like sun damage. The manual detection of damaged berries in the field is a subjective and labour-intensive task, and automatic detection by machine learning methods is challenging if all variants of damage should be modelled. Our proposed method detects regions of damaged berries in images in an efficient and objective manner using a shallow neural network, where the severeness of the damage is visualized with a heatmap. We compare the results of the shallow, fully trained network structure with an ImageNet-pretrained deep network and show that a simple network is sufficient to tackle our challenge. Our approach works on different grape varieties with different berry colours and is able to detect several cases of damaged berries like cracked berry skin, dried regions or colour variations.