Automatic detection of trapping events of postnatal piglets in loose housing pen: comparison of YOLO versions 4, 5, and 8
In recent years, the pig industry has experienced an alarming surge in piglet mortality shortly after farrowing due to crushing by the sow. This issue has been exacerbated by the adoption of hyperprolific sows and the transition to loose housing pens, adversely affecting both animal welfare and prod...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Korean Society of Animal Sciences and Technology
2025-05-01
|
| Series: | Journal of Animal Science and Technology |
| Subjects: | |
| Online Access: | http://www.ejast.org/archive/view_article?doi=10.5187/jast.2024.e106 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1850128544208781312 |
|---|---|
| author | Taeyong Yun Jinsul Kim Jinhyeon Yun Tai-Won Um |
| author_facet | Taeyong Yun Jinsul Kim Jinhyeon Yun Tai-Won Um |
| author_sort | Taeyong Yun |
| collection | DOAJ |
| description | In recent years, the pig industry has experienced an alarming surge in piglet mortality shortly after farrowing due to crushing by the sow. This issue has been exacerbated by the adoption of hyperprolific sows and the transition to loose housing pens, adversely affecting both animal welfare and productivity. In response to these challenges, researchers have progressively turned to artificial intelligence of things (AIoT) to address various issues within the livestock sector. The primary objective of this study was to conduct a comparative analysis of different versions of object detection algorithms, aiming to identify the optimal AIoT system for monitoring piglet crushing events based on performance and practicality. The methodology involved extracting relevant footage depicting instances of piglet crushing from recorded farrowing pen videos, which were subsequently condensed into 2-3 min edited clips. These clips were categorized into three classes: no trapping, trapping, and crushing. Data augmentation techniques, including rotation, flipping, and adjustments to saturation and contrast, were applied to enhance the dataset. This study employed three deep learning object recognition algorithms–– You Only Look Once (YOLO)v4-Tiny, YOLOv5s and YOLOv8s––followed by a performance analysis. The average precision (AP) for trapping detection across the models yielded values of 0.963 for YOLOv4-Tiny, and 0.995 for both YOLOv5s, and YOLOv8s. Notably, trapping detection performance was similar between YOLOv5s and YOLOv8s. However, YOLOv5s proved to be the best choice considering its model size of 13.6 MB compared to YOLOv4-Tiny’s 22.4 MB and YOLOv8’s 21.4 MB. Considering both performance metrics and model size, YOLOv5s emerges as the most suitable model for detecting trapping within an AIoT framework. Future endeavors may leverage this research to refine and expand the scope of AIoT applications in addressing challenges within the pig industry, ultimately contributing to advancements in both animal husbandry practices and technological solutions. |
| format | Article |
| id | doaj-art-f007a03f0f4142caa31496cbd274bfbf |
| institution | OA Journals |
| issn | 2672-0191 2055-0391 |
| language | English |
| publishDate | 2025-05-01 |
| publisher | Korean Society of Animal Sciences and Technology |
| record_format | Article |
| series | Journal of Animal Science and Technology |
| spelling | doaj-art-f007a03f0f4142caa31496cbd274bfbf2025-08-20T02:33:15ZengKorean Society of Animal Sciences and TechnologyJournal of Animal Science and Technology2672-01912055-03912025-05-0167366667610.5187/jast.2024.e106Automatic detection of trapping events of postnatal piglets in loose housing pen: comparison of YOLO versions 4, 5, and 8Taeyong Yun0Jinsul Kim1Jinhyeon Yun2Tai-Won Um3Department of Data Science, Chonnam National University, Gwangju 61186, KoreaSchool of Electronics and Computer Engineering, Chonnam National University, Gwangju 61186, KoreaDepartment of Data Science, Chonnam National University, Gwangju 61186, KoreaDepartment of Data Science, Chonnam National University, Gwangju 61186, KoreaIn recent years, the pig industry has experienced an alarming surge in piglet mortality shortly after farrowing due to crushing by the sow. This issue has been exacerbated by the adoption of hyperprolific sows and the transition to loose housing pens, adversely affecting both animal welfare and productivity. In response to these challenges, researchers have progressively turned to artificial intelligence of things (AIoT) to address various issues within the livestock sector. The primary objective of this study was to conduct a comparative analysis of different versions of object detection algorithms, aiming to identify the optimal AIoT system for monitoring piglet crushing events based on performance and practicality. The methodology involved extracting relevant footage depicting instances of piglet crushing from recorded farrowing pen videos, which were subsequently condensed into 2-3 min edited clips. These clips were categorized into three classes: no trapping, trapping, and crushing. Data augmentation techniques, including rotation, flipping, and adjustments to saturation and contrast, were applied to enhance the dataset. This study employed three deep learning object recognition algorithms–– You Only Look Once (YOLO)v4-Tiny, YOLOv5s and YOLOv8s––followed by a performance analysis. The average precision (AP) for trapping detection across the models yielded values of 0.963 for YOLOv4-Tiny, and 0.995 for both YOLOv5s, and YOLOv8s. Notably, trapping detection performance was similar between YOLOv5s and YOLOv8s. However, YOLOv5s proved to be the best choice considering its model size of 13.6 MB compared to YOLOv4-Tiny’s 22.4 MB and YOLOv8’s 21.4 MB. Considering both performance metrics and model size, YOLOv5s emerges as the most suitable model for detecting trapping within an AIoT framework. Future endeavors may leverage this research to refine and expand the scope of AIoT applications in addressing challenges within the pig industry, ultimately contributing to advancements in both animal husbandry practices and technological solutions. http://www.ejast.org/archive/view_article?doi=10.5187/jast.2024.e106Piglet crushingDeep learning object-detection algorithmYou Only Look Once (YOLO)TrappingArtificial intelligence of things (AIoT) |
| spellingShingle | Taeyong Yun Jinsul Kim Jinhyeon Yun Tai-Won Um Automatic detection of trapping events of postnatal piglets in loose housing pen: comparison of YOLO versions 4, 5, and 8 Journal of Animal Science and Technology Piglet crushing Deep learning object-detection algorithm You Only Look Once (YOLO) Trapping Artificial intelligence of things (AIoT) |
| title | Automatic detection of trapping events of postnatal piglets in loose housing pen: comparison of YOLO versions 4, 5, and 8 |
| title_full | Automatic detection of trapping events of postnatal piglets in loose housing pen: comparison of YOLO versions 4, 5, and 8 |
| title_fullStr | Automatic detection of trapping events of postnatal piglets in loose housing pen: comparison of YOLO versions 4, 5, and 8 |
| title_full_unstemmed | Automatic detection of trapping events of postnatal piglets in loose housing pen: comparison of YOLO versions 4, 5, and 8 |
| title_short | Automatic detection of trapping events of postnatal piglets in loose housing pen: comparison of YOLO versions 4, 5, and 8 |
| title_sort | automatic detection of trapping events of postnatal piglets in loose housing pen comparison of yolo versions 4 5 and 8 |
| topic | Piglet crushing Deep learning object-detection algorithm You Only Look Once (YOLO) Trapping Artificial intelligence of things (AIoT) |
| url | http://www.ejast.org/archive/view_article?doi=10.5187/jast.2024.e106 |
| work_keys_str_mv | AT taeyongyun automaticdetectionoftrappingeventsofpostnatalpigletsinloosehousingpencomparisonofyoloversions45and8 AT jinsulkim automaticdetectionoftrappingeventsofpostnatalpigletsinloosehousingpencomparisonofyoloversions45and8 AT jinhyeonyun automaticdetectionoftrappingeventsofpostnatalpigletsinloosehousingpencomparisonofyoloversions45and8 AT taiwonum automaticdetectionoftrappingeventsofpostnatalpigletsinloosehousingpencomparisonofyoloversions45and8 |