S18 - Session O4 - A Deep Learning-based Web Application for Segmentation and Quantification of Blueberry Internal Bruising

Authors: Xueping Ni, Fumiomi Takeda, Huanyu Jiang, Wei Yang, Seiya Saito, Changying Li *

Blueberries is an important fruit crop around the world. Recently, due to the high cost of labor for hand harvesting, many growers have turned to mechanical harvesters for harvesting their blueberries. Mechanical harvesters, however, cause fruit bruises which lead to substantial economic losses. Therefore, the blueberry bruise assessment is necessary to provide evidence to improve the harvesters in the field and efficiency of fruit sorting process in the packing house. Generally, packers check for fruit firmness of a few blueberries by means of tactile feel and with a firmness measuring instrument such as the FirmTech II. The packers are now assessing bruise damage in sliced fruit samples using visual scores, which is subjective and time-consuming. This study developed a deep learning-based web application that can automatically determine the blueberry bruise levels so the users can evaluate the mechanical harvesters in the field as well as packing lines. We annotated training images to train convolutional neural network models, obtaining models to detect and segment berries and bruised areas. The average precision (AP) for the berry detection model was 0.977 under 0.5 intersections of union (IoU) threshold for the validation dataset. The mean IoU for berry segmentation and bruise segmentation was 0.979 and 0.773, respectively for the validation dataset. The linear regression for the evaluated bruise results have a high correlation with the ground truth that was manually annotated. The results processed by three different hardware configurations were generated and compared with the ground truth. The developed website application is a robust tool for blueberry breeders and growers to evaluate berry internal bruises created by mechanical harvesters in the field.