⚡ OPENCV 5.0.0 • AWS GRAVITON NEOVERSE • FASTMCP • BEDROCK NOVA-LITE

Autonomous Active Vision Metrology
Eliminating Specular Glare on Metallic Foils

Specular reflection blinds traditional inspection systems on pharmaceutical blister packaging and battery pouch cells. SpecuLab-MCP combines sub-millimeter ArUco planar rectification, two-stage hysteresis glare gating, and Model Context Protocol (FastMCP) agentic loops to tilt viewpoints and fuse clear perspectives autonomously.

Launch Interactive Studio View GitHub Repository

5 / 7 (71.4%)

Strict Defect Recall

100% of resolvable defects ≥ 0.35 mm (vs 0/7 passive)

6 / 7 (85.7%)

Safety Capture Rate

Includes quarantine escalation for persistent glare

0 / 3 (0.0%)

Clean False Rejects

Eliminates 100% of passive glare false alarms

204.5 ms

Pipeline Latency

SIMD accelerated by Arm KleidiCV & AVX2 kernels

// INTERACTIVE MULTI-VIEW METROLOGY STUDIO

Experience Autonomous Glare Healing Live

Select an adversarial held-out specimen below. Compare the blinded primary camera frame against our USAC-MAGSAC homography fused active canvas.

↔

Ground Truth Target 0.95 mm Puncture at (950, 400)

Residual Glare % 0.0%

Active Views Fused View 2 (View 1 Rejected for 94% Overlap)

Inspection Verdict

REJECT — Defect Confirmed

// EMPIRICAL 6-METHOD BENCHMARK

Defeating the Glare Barrier Across 10 Held-Out Cards

Evaluated on our 14-specimen Automated Procedural Optical Testbed under realistic twin-lobe LED striations, anisotropic brushed foil, and multi-angle perspective tilts.

Inspection Method	Strict Recall	Safety Capture (incl. Escalate)	Clean False Rejects	Unresolved Occlusions
`1A_passive_unmasked`	1/7 (14.3%)	1/7 (14.3%)	3/3 (100% False Alarms)	0/10
`1B_passive_glare_masked`	0/7 (0.0% Blind)	0/7 (0.0% Blind)	0/3 (0.0%)	0/10
`2A_naive_view1`	2/7 (28.6%)	3/7 (42.9%)	0/3 (0.0%)	1/10
`2B_greedy_lowest_glare`	3/7 (42.9%)	4/7 (57.1%)	0/3 (0.0%)	1/10
`2C_active_deterministic`	5/7 (71.4%)*	6/7 (85.7%)*	0/3 (0.0%)	1/10
`2D_active_bedrock_live`	5/7 (71.4%)*	6/7 (85.7%)*	0/3 (0.0%)	1/10

*Note: 5/7 strict recall represents 5/5 (100.0%) of all physically resolvable defects ≥ 0.35 mm. Specimen 11 contains an intentional 0.22 mm sub-threshold micro-dimple below optical resolution limits, honestly documented across all methods.

// AGENTIC VISION & MODEL CONTEXT PROTOCOL

FastMCP Atomic Tool Execution Pipeline

Amazon Bedrock (Nova-Lite) dynamically drives the vision perception-decision-action loop using 3 atomic, stateless FastMCP tools with strict schema validation.

Specimen Decision Timeline

FastMCP Tool Interface

Stateless protocol registered at src/mcp_server.py:

// 1. rank_candidate_views
{
  "name": "rank_candidate_views",
  "inputSchema": {
    "primary_image_path": "string (filepath)",
    "candidate_image_paths": ["string (filepath)"]
  },
  "returns": {
    "ranked_candidates": [
      {
        "candidate_path": "assets/specimens/specimen_06/view_2.jpg",
        "glare_percentage": 6.5,
        "overlap_percentage_with_primary": 0.0,
        "relative_sharpness_ratio": 1.05,
        "is_blurry": false,
        "multi_criteria_rank": 1
      }
    ]
  }
}

// 2. fuse_candidate_view
// Applies USAC_MAGSAC, radiometric gain matching, 
// and 11 px feather blending. Max fusions = 2.

// 3. finalize_inspection_or_escalate
// Detects defects via Blackhat 61x61 morphological closing.
// Halts line with ESCALATE if glare > 2.0% threshold.

// AUTHENTIC MICRO-BENCHMARKS & ARM ACCELERATION

Hardware SIMD Kernel Profiling

1-Thread SIMD Isolation: Evaluates single-core vector efficiency without multi-core scheduling noise.

Empirical Stage Latency & Speedup

OpenCV 5.0.0 SIMD Optimized vs Unoptimized Baseline

Hardware Label Honesty

Committed JSON Provenance

Local Baseline Host:
Windows-11-10.0.26300-SP0 (AMD64) running OpenCV 5.0.0 AVX2 SIMD.

Automated AWS Graviton Neoverse CI:
Runs on real 64-bit Arm (ubuntu-24.04-arm aarch64) in GitHub Actions.

View Verified CI Run

// PHYSICAL TESTBED & ARUCO CALIBRATION

Sub-Millimeter Physical Calibration Kit

Download and print the canonical dual-ArUco card at 100% scale (no fit-to-page) to replicate active multi-view metrology in your physical laboratory.

Canonical Pixel & Physical Specs

Active Card Canvas: 100.0 mm x 40.0 mm
Resolution: 2000 px x 800 px (Nominal 20.0 px/mm scale)
Primary Marker: ArUco ID 0 (400x400 px, 20x20 mm) at (100, 200)
Secondary Marker: ArUco ID 1 (400x400 px, 20x20 mm) at (1500, 200)
Central Foil ROI: 40.0 mm x 20.0 mm (800x400 px)
ORB Texture Bands: Deterministic high-frequency gradient bars

Download High-Res Card PNG (2000x800)

Autonomous Active Vision Metrology Eliminating Specular Glare on Metallic Foils