Reliability-aware query restoration for embedded real-time object detection in degraded underwater vision systems

Embedded underwater vision systems require real-time object detection that remains reliable in the face of turbidity, color attenuation, low contrast, weak boundaries, and cluttered backgrounds. In real-time detection transformer (RT-DETR)-style detectors, these degradations affect the ranking and selection of encoder candidates used to initialize decoder queries. This study proposes a degradation-aware query reliability restoration framework for RT-DETR, termed DQR-RTDETR. The Candidate Reliability Prior (CRP) estimates candidate trustworthiness from encoder memory, Stability-Preserving Query Reliability Recalibration (QRR) adjusts top-K candidate scores and selected-query features, and Scale-Aware Local Evidence Compensation (SLEC) preserves local cues for small or weakly structured targets. On the SeaClear Marine Debris Detection and Segmentation Dataset (SeaClear), averaged across three random seeds, DQR-RTDETR improves mean average precision across intersection-over-union thresholds from 0.50 to 0.95 (mAP@0.5:0.95) from 0.6998 ± 0.0029 to 0.7444 ± 0.0026 and average precision at an intersection-over-union threshold of 0.75 (AP75) from 0.8030 ± 0.0038 to 0.8615 ± 0.0044, while adding only 0.05 million parameters. Validation on the TrashCan underwater marine-debris dataset and the Detecting Underwater Objects dataset (DUO), degradation-grouped SeaClear analysis, and query-level diagnostics show consistent gains in accuracy, strict localization, and selected-query quality. Deployment on an NVIDIA Jetson Orin NX using Robot Operating System 2 (ROS 2) and TensorRT achieves 18.7 frames per second (FPS) with a mean end-to-end latency of 52.8 milliseconds. These results indicate that restoring query reliability is an effective, compact intervention for edge-deployable underwater electronic vision systems.