Publication: Ginger seed growth recognition using mask region based convolutional neural network (mask r-cnn)
No Thumbnail Available
Date
2023-01-01
Authors
Tong Yin Syuen
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
As a plant that poses unique culinary and medical uses, ginger has emerged as
a valuable commodity in Asia. Among the critical processes in the production of ginger
is ginger seed preparation. It is particularly important to monitor the growth and quality
of ginger seeds before they are being sown in growing media to ensure germination.
However, to date, the ginger seed monitoring process remains manual and is reliant on
human experts, despite the growing demand for more effective and accurate
monitoring. In this work, a total 1,746 images consisting 2,230 sprout instances were
collected from 282 ginger seed samples. In order to realize the automatic monitoring
of ginger seeds, deep learning architectures were employed to detect the ginger seed
sprouts in three stages from the digital images. This work assessed and compared the
instance segmentation task using end-to-end Mask R-CNN models built by different
strategies. Then, a two-stage hybrid detector-classifier model was also proposed to
benefit from model task specialization concept. Specifically, an end-to-end binaryclass
Mask R-CNN and multi-class classifier were combined to be compared to an
end-to-end multi-class Mask R-CNN. The experimental results indicate that the use of
the hybrid detector-classifier model developed in this work achieved mAP0.50 of
84.27% at inference time of 0.383 second per image in the detection of 402 images
consisting of 514 sprout instances. Besides, substantial confusion between object
classes in the model was also observed to be in line with the human expert’s perception
in data annotation.