Citation

Hasan, Md Najmul (2026) Chicken health monitoring through analysis of droppings using machine learning. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

In this study, the contribution of the poultry industry towards Malaysia’s food security is investigated, with open-house chicken farms portrayed as the dominant mode of production. Health monitoring of the chickens of these farms is still carried out manually, laborious and risky as it is based on the common eyes on the droppings of the chickens. Because many poultry diseases occur earlier on by analyzing the change in droppings, the thesis segment assumes the recognition of chicken health status through droppings analysis using machine learning in Malaysian open-house farms. A dataset was constructed under real open-house Malaysian farm conditions, capturing challenges such as cluttered wooden slat flooring, irregular dropping shapes, and variable lighting. A polygon annotation tool was used to annotate the dataset with precise shape of droppings, while advanced augmentation methods, such as Mosaic4, cut-out, rotation, blur, brightness and contrast change, and noise injection, increased the diversity of the data and solved the problem of imbalance in the classes. The interclass discrimination was significantly enhanced by Mosaic4 using multi-class composite images. Seven lightweight detectors: MobileNetV3-SSD, EfficientDetLite0, YOLOv5n, YOLOv7-tiny, YOLOv8n, YOLOv10n, and YOLO11n were benchmarked using mAP, precision, recall, speed, and efficiency. YOLO11n provided the best baseline trade-off, achieving mAP@0.5:0.95 of 83.50% with 1.20 ms/image inference and 2.58M parameters, outperforming MobileNetV3-SSD and EfficientDetLite0 under the same farm constraints. The baseline YOLO comparison further confirmed YOLO11n as both highly accurate and among the fastest models tested. Building on this baseline, the framework was optimized through targeted architectural and training improvements, producing a final model integrating C3Ghost (neck), SimAM (attention), SiLU (activation), and tuned hyperparameters. The resulting optimized YOLO11n achieved mAP@0.5 of 99.2% and mAP@0.5:0.95 of 88.5% while remaining lightweight with 2.29M parameters, 5.8 GFLOPs, and competitive inference speed of 2.3 ms/image. This thesis contributes a Malaysia-specific openhouse chicken droppings dataset and an efficient, localization-capable droppings detection model, providing a strong foundation for practical early-warning chicken health monitoring systems in resource-constrained farm settings.

Item Type:	Thesis (Masters)
Additional Information:	Call No.: Q325.5 .M36 2026
Uncontrolled Keywords:	Machine learning—Technological innovations
Subjects:	Q Science > Q Science (General) > Q300-390 Cybernetics
Divisions:	Faculty of Artificial Intelligence & Engineering (FAIE)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	22 May 2026 04:28
Last Modified:	22 May 2026 04:28
URII:	http://shdl.mmu.edu.my/id/eprint/15898

Downloads

Edit (login required)