Static Single-Pass Baseline
Fails under degradation
Green = Ground Truth Defect | Red = Baseline False/Missed Box
Closed-Loop Agentic OpenCV 5
Converged
Green = Ground Truth | Amber = Verified Agentic Detection
Live Perception Telemetry
9.65 ms (COOL Arm64)Agentic Tool Execution Chain (MCP)
Ready to export telemetry
60-Scene Agentic F1 Score
0.9899
+0.6337 over static baseline (0.3562)
Precision / False Positives
100.0%
0 false alarms across all 60 scenes
AWS Graviton4 COOL Speedup
2.60x
9.65 ms (103.6 FPS) on c8g.xlarge
Cloud Inspection Cost Savings
-65.9%
$0.49 vs $1.44 per 1M frames
Reproducible 60-Scene Benchmark & AWS Graviton4 COOL Kernel Breakdown
| Degradation Regime / Kernel | Static Baseline | Closed-Loop Agentic OpenCV | x86_64 Latency | Graviton4 COOL (Arm64) |
|---|---|---|---|---|
| Specular Glare Domes (`glare`) | 0.8000 F1 | 1.0000 F1 (+0.2000) | 19.4 ms | 7.4 ms (2.62x) |
| Underexposed Low-Contrast (`low_contrast`) | 0.0000 F1 | 1.0000 F1 (+1.0000) | 21.2 ms | 8.2 ms (2.59x) |
| Defocus / Motion Blur (`defocus_blur`) | 0.1429 F1 | 1.0000 F1 (+0.8571) | 24.8 ms | 9.5 ms (2.61x) |
| Compound Glare + Blur + Low Contrast | 0.0000 F1 | 0.9577 F1 (+0.9577) | 25.1 ms | 9.65 ms (2.60x) |
Model Context Protocol (MCP) & Amazon Bedrock AgentCore Integration
Unlike static vision wrappers where an LLM merely describes a fixed detector output, our system uses OpenCV 5 quantitative perception metrics (`laplacian_sharpness`, `glare_fraction`, `mean_luminance`, `contrast_std`) to actively alter subsequent OpenCV tool invocations, kernel sizes, and AWS IoT Core actuation decisions.