Comparative analysis of deep learning models for sheep detection in aerial imagery

• 主要な著者: Ismail, Muhammad Syahmie, Samad, Rosdiyana, Pebrianti, Dwi, Mustafa, Mahfuzah, Abdullah, Nor Rul Hasma
• フォーマット: Proceeding Paper
• 言語: English; English
• 出版事項: IEEE 2024
• 主題: T Technology (General)
• オンライン･アクセス: http://irep.iium.edu.my/115359/7/115359_Comparative_Analysis_of_Deep_Learning_Models_for_Sheep_Detection_in_Aerial_Imagery.pdf; http://irep.iium.edu.my/115359/13/115359_Comparative%20Analysis%20of%20Deep%20Learning%20Models%20for%20Sheep%20Detection%20in%20Aerial%20Imagery_Scopus.pdf; http://irep.iium.edu.my/115359/; https://ieeexplore-ieee-org.ezlib.iium.edu.my/document/10652292

This research evaluates You Look Only Once - YOLOv5, YOLO-NAS, and Detection Transformer (DETR) and provides a thorough evaluation of deep learning models for sheep identification in aerial pictures. A carefully selected collection of 4,212 aerial photos of sheep in various environments was used to thoroughly evaluate model performance. The implementation involved preprocessing, augmentation, model parameter optimization, training on Google Collab GPU s, and quantitative test results analysis. Important results show that on the sheep dataset, YOLOv5 and YOLO-NAS achieved an impressive accuracy of 97%, exceeding DETR's initial accuracy range of 70-80%. However, after adjusting the hyperparameters, DETR's accuracy significantly increased to 86%, showing less overfitting and more stability. The increased accuracy of YOLO models highlights how useful they are for sheep counting and aerial surveillance to support modern farming techniques. However, improvements to the transformer based DETR may increase its usefulness even more. This research offers valuable insights into the real-world applications of deep learning for livestock detection in aerial imagery, providing a foundation for future advancements in the field.