Deep learning based image segmentation for grape bunch detection

This paper presents a method for automatic fruit detection in vineyards through the inspection of color images obtained by a low cost RGB-D sensor placed onboard an agricultural vehicle. Image segmentation is obtained by using a pre-trained convolutional neural network, which receives input data, as sub-patches of known size, and performs the classification in few classes of interest. Output scores are then used to create the probability maps for each class and, thus, pixel-by-pixel segmentation of the grape clusters. Field experiments prove the ability of the proposed processing to successfully segment grape clusters, with accuracy of 87.5%, despite the poor quality of the input images.