Generalization of deep learning models applied to semantic segmentation of in-field natural images in vineyards

Extracting knowledge from natural images captured in-field requires semantic segmentation to precisely detect fruits or other objects of interest. This paper demonstrates the generalizability of deep networks trained on a highly variable dataset in the context of grape segmentation in vineyards. Among the proposed architectures, the DeepLabv3-ResNext returned the best results, scoring a noticeable 95.1% balanced accuracy during the test. This best model was then transferred to segment images of other datasets, captured by different cameras, and focusing on other grape varieties. Results proved the robustness of the DeepLabv3-ResNext, thus bringing the proposed methodology closer to its actual use.