Probability map-based grape detection and counting

Abstract

Grape-cluster detection, grape-berry counting, and berry-per-grape-cluster counting are the three major tasks in digital viticulture. Although the detection of clusters and berries was treated as separate tasks in previous studies, due to the clustered nature of the grape phenotype and the close hierarchical spatial distribution relationship between grape clusters and berries, those three tasks should be considered at the same time in our view; further, we define these tasks as subtasks of the grape detection and counting task. The major challenge of combining the three subtasks of grape detection and counting together is how to obtain the detection results of those subtasks from the output of a single neural network. To address this issue, we propose a Probability Map-based Grape Detection and Counting framework that first detects two intermediate probability maps through a neural network and utilizes three postprocessing stages to finish the three grape detection and counting subtasks sequentially. After being trained and validated on the trainval set of WGISD, combining 100 extra images collected in Chengdu by our team, the proposed framework is tested on the test set of WGISD to verify its performance. The proposed framework achieves a localization performance of average precision (AP)(iou0.5) 0.851 and a counting performance of MAE 1.845, RMSE 2.142 for grape clusters, a counting performance of MAE 23.414, RMSE 31.391 for grape berries, and a counting performance of MRD 0.142, 1-FVU 0.865 for berries per grape cluster on the test set of WGISD. Our source code will be publicly available athttps://github.com/volcanoYcc/Probability-Map-based-Grape-Detection-and-Counting.