An IoT-integrated YOLOv8 framework with adaptive domain alignment for real-time quality inspection in smart logistics

Abstract This paper presents an IoT-integrated YOLOv8 framework for real-time quality inspection in smart logistics. The system reformulates industrial anomaly detection as supervised, logistics-specific defect detection with actionable bounding-box outputs that can be directly consumed by warehouse management systems (WMS). On the vision side, we introduce a small-defect–biased detection head and class-balanced loss weighting to improve sensitivity to rare, fine-grained defects while controlling false alarms. A lightweight style-transfer–based domain adaptation module further narrows the visual gap between industrial anomaly datasets and warehouse imagery, improving robustness to illumination changes, label gloss, and reflective packaging. On the IoT side, an MQTT-based architecture publishes defect alerts from edge devices to WMS dashboards and implements an adaptive feedback policy that triggers incremental retraining when rolling error rates exceed a predefined threshold. Experiments on MVTec AD yield 97.6% [email protected] at 72 FPS, while cross-dataset evaluations on VisA and MVTec AD 2 incur only a 3.8% relative drop in mAP. Under controlled logistics-inspired test conditions involving illumination variability and reflective packaging, the IoT-integrated prototype achieved a stable end-to-end response latency of 20–30 ms and MQTT message-delivery reliability exceeding 98% during a continuous 48-h evaluation. The adaptive feedback mechanism further supports robustness to operational drift through periodic incremental retraining.