End-to-End Automatic Berry Counting for Table Grape Thinning

Bunch shape and berry size indicate the quality of table grapes and crucially affect their market value. Berry thinning is one of the most important tasks in grape cultivation to achieve an ideal bunch shape and to make sufficient space for individual berries. A successful practice by skilled grape farmers in Japan is using the number of berries in a bunch to guide the thinning process; hence, a technique for automatically counting the number of berries in a working bunch has been long desired by farmers to improve the efficiency of the thinning task. This research presents a novel end-to-end berry-counting technique based on a deep neural network (DNN), and its contributions are as follows. First, because a DNN requires massive training data, a novel data augmentation technique simulating the thinning process is proposed. Second, a new location-sensitive object detection model that integrates explicit location information and supplementary classification loss into a state-of-the-art instance segmentation model was proposed for detecting the number of berries in a working bunch with a high accuracy. Third, a set of features, together with their extraction algorithms, is designed for predicting the number of berries in a bunch (3D counting) using the berries detected on a single 2D image. Experiments using data collected from farmers' grape-thinning process have been conducted to validate the accuracy and effectiveness of the proposed methods.

Automated summary

This research introduces a novel end-to-end berry-counting technique based on a deep neural network, which utilizes a new data augmentation technique and object detection model to achieve accurate prediction of the number of berries in a working bunch.