Edge AI-enabled chicken health detection based on enhanced FCOS-Lite and knowledge distillation

Abstract

Edge-AI based AIoT technology offers significant benefits advantages in modern poultry management by optimizing farming operations and reducing resource requirements. To address the challenge of developing a highly accurate and lightweight edge-AI enabled detector that can be deployed within memory-constrained edge environments, this study propose an innovative real-time, compact and highly accurate edge-AI enabled detector, based on improved FCOS-Lite and designed to detect chickens and their health status using a highly resource-constrained edge-AI enabled CMOS sensor. The proposed FCOS-Lite detector leverages MobileNet as the backbone to achieve a compact model size. To mitigate the issue of reduced accuracy in compact edge-AI detectors without incurring additional inference costs, we propose a gradient weighting loss function for classification and introduce a CIOU loss function for localization. Furthermore, a knowledge distillation scheme is employed to transfer critical information from a larger teacher detector to the FCOS-Lite detector, enhancing performance while preserving the compactness. Experimental results demonstrate the proposed detector achieves a mean average precision (mAP) of 95.1% and an F1-score of 94.2%, outperforming other state-of-the-art detectors. The detector operates efficiently at over 20 FPS on the edge-AI enabled CMOS sensor, enabled by int8 quantization. These results confirm that the proposed innovative approach leveraging edge-AI technology achieves high performance and efficiency in a memory-constrained environment, meeting the practical demands of automated poultry health monitoring with low power consumption and minimal bandwidth costs.