Towards practical 2D grapevine bud detection with fully convolutional networks

In Viticulture, visual inspection of the plant is a necessary task for measuring relevant variables. many cases, these inspections are susceptible to automation through computer vision methods. Bud detection one such task, central measurement important variables as: bud sunlight exposure, autonomous pruning, counting, type-of-bud classification, geometric characterization, internode length, area, and development stage, among others. This paper presents method grapevine based on Fully Convolutional Networks MobileNet architecture (FCN-MN). To validate its performance, this was compared in with strong detection, Scanning Windows (SW) patch classifier, showing improvements over three aspects detection: segmentation, correspondence identification localization. The best version FCN-MN showed F1-measure $88.6\%$ (for true positives defined as detected components whose intersection-over-union above $0.5$), false that small near bud. Splits -- overlapping mean segmentation precision $89.3\% (21.7)$, while alarms not pixel area only $8\%$ bud, distance (between mass centers) $1.1$ diameters. concludes by discussing how results would produce sufficiently accurate measurements number, suggesting good performance practical setup.