YOLO-v5를 이용한 등검은말벌집 탐색 기술 연구 (Study on the Technology for Searching Vespa Velutina Nest Using YOLO-v5)

서지정보

· 발행기관 : 한국양봉학회
· 수록지 정보 : Journal of Apiculture / 37권 / 3호 / 255 ~ 263페이지
· 저자명 : 정유석, 전문석, 김수배, 김동원, 유승화, 김경철, 이시영, 이창우, 최인찬

초록

Vespa velutina, such as ecological disturbance wildlife introduced into Busan through Shanghai, China in 2003, is an ecologicla disturbance wildlife and difficult to find with the naked eye due to the nature of creating wasp nests in high places in the forest. In addition, its population increases every year, causing the economical damage to beekiipers (about 170 billion won).
Therefore, the purpose of this study is to acquire images and explore the location of Vespa velutina based on artificial intelligence (AI) to easily remove Vespa velutina using drones.
Differently from the ground images, aerial images have the characteristic that the objects become very small as the altitude rises, so it is necessary to study the appropriate image size that can be mounted on drones and can search Vespa velutina with AI. The AI model YOLO-v5 using four image sizes (640×384 px, 1,280×736, 1,920×1,088, 3,840×2,176) was applied in this study and the original size of the image (3,840×2,160) was used for learining and verification.
When confidence was higher than 0.7, F1 score of YOLO-v5s-default (640) learned with image size 640×384 was 2.4%, YOLO-v5s-1280 was 36.5%, YOLO-v5s-1920 was 64.2% and YOLOv5s- 3840 obtained the best detection performance at 96.1%. In addition, the performance was confirmed when the verification image size was 3,840×2,176 in four AI learned with different image sizes. In this case, YOLO-v5s-default (640) had F1 score of 4.4%, YOLO-v5s-1280 was 7.1% and YOLO-v5s-1920 was 16.4% with no performance improvement. Therefore, this study found that the networks that are learned and validated using the aerial images larger than 640 showed the better performance. Continuous research on real-time information sharing and estimation of the location of the occurrence of Vespa velutina through data collection using drone systems will be carried out.

영어초록

Vespa velutina, such as ecological disturbance wildlife introduced into Busan through Shanghai, China in 2003, is an ecologicla disturbance wildlife and difficult to find with the naked eye due to the nature of creating wasp nests in high places in the forest. In addition, its population increases every year, causing the economical damage to beekiipers (about 170 billion won).
Therefore, the purpose of this study is to acquire images and explore the location of Vespa velutina based on artificial intelligence (AI) to easily remove Vespa velutina using drones.
Differently from the ground images, aerial images have the characteristic that the objects become very small as the altitude rises, so it is necessary to study the appropriate image size that can be mounted on drones and can search Vespa velutina with AI. The AI model YOLO-v5 using four image sizes (640×384 px, 1,280×736, 1,920×1,088, 3,840×2,176) was applied in this study and the original size of the image (3,840×2,160) was used for learining and verification.
When confidence was higher than 0.7, F1 score of YOLO-v5s-default (640) learned with image size 640×384 was 2.4%, YOLO-v5s-1280 was 36.5%, YOLO-v5s-1920 was 64.2% and YOLOv5s- 3840 obtained the best detection performance at 96.1%. In addition, the performance was confirmed when the verification image size was 3,840×2,176 in four AI learned with different image sizes. In this case, YOLO-v5s-default (640) had F1 score of 4.4%, YOLO-v5s-1280 was 7.1% and YOLO-v5s-1920 was 16.4% with no performance improvement. Therefore, this study found that the networks that are learned and validated using the aerial images larger than 640 showed the better performance. Continuous research on real-time information sharing and estimation of the location of the occurrence of Vespa velutina through data collection using drone systems will be carried out.

참고자료

· 없음