The study developed a lightweight AI detection system, VetStar, enabling rapid and accurate antibiotic residue detection on embedded devices, significantly improving detection efficiency.
Literature Overview
This paper, 'AI-Powered Embedded System for Rapid Detection of Veterinary Antibiotic Residues in Food-Producing Animals', published in Antibiotics, reviews current research advancements and challenges in veterinary antibiotic residue detection. It highlights that traditional colloidal gold immunochromatography (CGIA) methods, while operationally simple, rely on manual interpretation leading to low efficiency and poor consistency. Existing detection algorithms suffer from slow inference speeds on embedded devices, failing to meet high-throughput requirements. The authors propose VetStar, a lightweight object detection algorithm combined with BCKD knowledge distillation, achieving rapid inference and high detection accuracy for food safety applications.
Background Knowledge
Veterinary antibiotics are widely used in animal husbandry but pose potential health risks including antimicrobial resistance, toxic reactions, and gut microbiota disruption. Colloidal gold immunochromatography (CGIA) remains the standard rapid screening method, yet its visual interpretation is susceptible to lighting conditions. Recent efforts have incorporated image processing and AI technologies to enhance efficiency and consistency, though most solutions still depend on cloud servers, limiting applicability in resource-constrained environments. While algorithms like YOLO strike a balance between accuracy and speed, their high parameter counts and computational demands hinder deployment on embedded systems. This study focuses on developing lightweight models for offline, high-throughput detection to support real-time monitoring in remote laboratories.
Research Methods and Experiments
The research team constructed an embedded detection system based on the Rockchip RK3568 platform, integrating a 5MP OV5640 auto-focus USB camera and COB LED illumination. The system automates image acquisition and AI interpretation, generating standardized detection reports to enhance detection standardization. The proposed VetStar algorithm features a lightweight feature extractor (StarBlock) and depthwise separable-reparameterized detection head (DR-head), substantially reducing parameters and computational load while maintaining accuracy. The study employed BCKD (Bridging Cross-task Protocol Inconsistency Knowledge Distillation) training, enabling lightweight student models to learn refined feature representations from teacher models and improve detection precision.
Key Conclusions and Perspectives
Research Significance and Prospects
This study presents a portable, high-throughput antibiotic residue detection system suitable for grassroots food safety laboratories, offering excellent real-time performance and interpretation consistency. It provides an efficient, cost-effective solution for veterinary antibiotic residue monitoring. Future work will focus on optimizing model architectures, expanding detection capabilities to other veterinary drug residues, and integrating advanced AI algorithms for multi-task detection to enhance system versatility and intelligence.
Conclusion
The study successfully developed a lightweight AI detection system, VetStar, combining the StarBlock feature extraction module with DR-head architecture to enable rapid inference on RK3568 embedded devices. The system demonstrated exceptional performance on the VDR-RTC dataset, achieving 97.4% mAP50 accuracy with only 5.4 seconds of inference time per sample - significantly outperforming YOLO series models. The BCKD knowledge distillation approach further improved precision while maintaining minimal resource consumption, ensuring reliable detection in resource-limited environments. This system holds promise for grassroots food safety screening, veterinary drug residue monitoring, and field rapid testing, advancing intelligent automation in food safety regulation.
