Application of mask region-based convolutional neural network on asparagus growth identification
Journal
American Society of Agricultural and Biological Engineers Annual International Meeting, ASABE 2021
Journal Volume
4
Pages
2180-2185
Date Issued
2021
Author(s)
Abstract
Asparagus (Asparagus officinalis L.) cultivation in Taiwan adopts the mother stalk method to boost production. This method operates under a controlled number of mother stalks to improve the growth rate of spears. However, extra effort on field maintenance results in labor costs. An automatic growth identification system was proposed to estimate the number of stalks and the length of spears using a deep neural network algorithm. The dataset was composed of more than 600 images of greenhouse asparagus with semantic annotation for three classes, spear, mother stalk, and clump. The three classes represented young and harvestable stems, old and well-grown stems, and gathered stalks, respectively. Mask region-based convolutional neural network (Mask R-CNN) with ResNeXt-101 as backbone was selected as the detection model. The number of stalks and the length of spears could be calculated based on the predicted bounding boxes and the masks from the model. The preliminary results showed average precisions (APs) of the spear, stalk, and clump being 79.34%, 80.23%, and 90.84% for localization; 73.71%, 60.59%, and 90.84% for segmentation. The estimated number of stalks was 4.57 averagely with a root mean square error (RMSE) of 1.06, and the estimated length of spear achieved a determination of coefficients (R2) of 0.8849. Web application for growth identification has been developed for practical use. ? American Society of Agricultural and Biological Engineers Annual International Meeting, ASABE 2021. All Rights Reserved.
Subjects
Asparagus
Computer vision
Deep neural network
Web application
Convolution
Cultivation
Deep neural networks
Mean square error
Semantics
Wages
Bounding box
Detection models
Neural network algorithm
Practical use
Region-based
Root mean square errors
Semantic annotations
WEB application
Convolutional neural networks
SDGs
Type
conference paper