Quantization-Aware Training for Minority-ClassPreserving Poultry Disease Detection on Edge Devices

Poultry diseases, notably Newcastle Disease (NCD), Coccidiosis, and Salmonellosis, have long caused significant economic losses in developing regions. Despite this, reliable diagnostic tools for resourceconstrained field environments remain limited. While deep learning models excel in labs, their accuracy drops sharply when compressed via PostTraining Quantization (PTQ) for edge devices, especially in minority classes. This study investigates PTQ performance degradation on MobileNetV3 Small using a four-class poultry feces dataset. Results show the NCD F1-score fell by 74.3% relatively (0.9444 to 0.2424), which is unacceptable for veterinary diagnostics. We propose a Quantization-Aware Training (QAT) workflow using PyTorch FX Graph Mode and the qnnpack backend. This restored the Macro F1 to 0.9522 and NCD F1 to 0.8621. The final 1.84 MB TFLite model achieved a 0.9714 Macro F1 score, with inference times of 2–30 ms on Android devices, demonstrating its feasibility for resource-limited edge hardware.