Identifying veraison process of colored wine grapes in field conditions combining deep learning and image analysis

Accurate identification of the veraison process is essential for improving wine quality, which is challenging due to the variability of veraison among berries of the same cluster in algorihtm design, and also the subjective and labor-intensive issues in mannual identification. Therefore, this study proposed a method combining deep learning and image analysis to identify veraison in colored
wine grapes under natural field growing conditions. The removal of irrelevant background was first achieved by semantic segmentation model, and then Mask R-CNN instance segmentation pipeline was constructed with anchor parameters optimization. In particular, three kinds of backbone networks were analyzed and compared in Mask R-CNN, and the overall performance of ResNet50-FPN was the best, with the testset Average Precision reaching 81.53% and the inference time being only 45.70ms/frame. Then, a method for characterizing berry veraison by H component of HSV color space was proposed and the invariance of the H component of three colored wine grape berries under different light conditions was verified and discussed. An algorithm was developed to identify
veraison progress by calculating the percentage of the number of berries of different grades in the total number of berries of the whole grape bunches. The test accuracy reached 92.50%, 91.25% and 91.88% for three wine grapes including Cabernet Sauvignon, Matheran and Syrah respectively. The proposed method is able to provide vital reference for automated monitoring and intelligent management decisions of grape growth.