大島商船高等専門学校紀要

Back to Top

大島商船高等専門学校紀要 Volume 54
published_at 2021-12

Investigating the effect of shape similarity on object detection

物体検出における形状の類似性が与える影響の調査
Harada Masafumi
fulltext
4.51 MB
OS10054000004.pdf
Descriptions
In recent years, various damages by crow, a verminous bird has increased. Our research group have been developing an automatic eviction system for verminous birds using drones to reduce the damage caused by crow, cormorants, etc. The input to this system is images of verminous birds taken by a camera mounted on a drone. This image will be used as training data for Convolutional Neural Network. However, the problem with this method is that it cannot identify birds that are similar in color and shape. Therefore, this paper focuses on the shape information, and training has been provided by YOLOv3. As a result, an airplane similar in shape to a crow couldn’t be identified as a crow properly. From this, it was confirmed that shape similarity is an impediment to identification of birds. So, training that emphasizes the characteristics of each bird species is necessary.
Creator Keywords
Convolutional Neural Network(CNN)
You Only Look Once(YOLO)
Data Augmentation(DA)