首页|A deep learning-based web application for segmentation and quantification of blueberry internal bruising
A deep learning-based web application for segmentation and quantification of blueberry internal bruising
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
Elsevier
? 2022 Elsevier B.V.Blueberries have become an important fruit crop around the world. Most blueberries are hand-harvested to maintain high quality before the packing process for fresh market distribution. However, in the last 10 years increasing acreage has been machine harvested to reduce labor costs although machine harvesting causes bruising in more than 20% of blueberries. Non-bruised blueberries remain firmer and can be cold stored longer than bruised blueberries, while the bruised fruit cannot be sold, leading to a substantial economic loss. Current packing line sorting technology can sort soft berries but is unable to detect and sort out bruised fruit. Blueberry bruise assessment is necessary for providing the means to improve the harvester efficiency and fruit sorting process in the packing house. The goal of this study was to develop a web browser-based application (Web App) that users can access easily and determine the blueberry bruises accurately and quickly. We annotated 1725 blueberries to train MobileNet SSD and MobileNet-UNet, two deep learning models, generating a berry detection model, a berry segmentation model, and a bruise segmentation model. The average precision (AP) for the berry detection model was 0.977. The mean intersection over union (IoU) for berry segmentation and bruise segmentation was 0.979 and 0.773, respectively. Bruises were assessed for 56 images with 50 sliced blueberries in each image using the trained models that implemented in the Web App we developed. The bruise ratio data obtained from three different hardware devices were compared with the results that were manually annotated. Linear regression analyses showed a high correlation between the results from the deep learning models and the ground truth. The bruise ratio prediction using three hardware devices achieved an accuracy of 78.7%, 79.0%, and 78.9%, respectively, indicating that the model performance was satisfactory regardless of the hardware configuration. The average processing time for each image under three hardware configurations revealed that our Web App was superior to the manual method. The Web App reduced the time needed for assessing and tabulating bruise damage for 50 fruit samples from about 15 min needed for manual visual method to less than 30 s with comparable accuracy. This Web App is a robust tool for blueberry breeders, farmers and packers for evaluating berry bruises.
NSTL
Elsevier
