🌾 OPTIX : Adaptive Hyperspectral Inspection System

Tagline: Scan • Suppress • Decide • Isolate • Retrain

Smart India Hackathon (SIH) 2026 · Problem Statement 26233
Inline Microbial Contamination Detection Using Hyperspectral Edge Sensors
Theme: Agriculture, FoodTech & Rural Development · Category: Hardware
Team Name: Team AEROS (Team ID: 186870)
Live Interactive Simulator: https://sandy1472006.github.io/OPTIX/

---

📌 Executive Summary

Foodborne microbial contamination affects over 600 million people worldwide annually. In modern food processing facilities (poultry processing, meat packing, and fresh leafy greens washing lines), conventional quality assurance relies on destructive, random spot sampling:

• Laboratory Culture Testing: Accurate, but requires $24 - 48\text{ hours}$ of incubation—by the time contamination is discovered, entire product batches have already entered the retail cold chain.
• PCR & Molecular Testing: Highly specific, but slow, expensive, and limited to minute off-line swab aliquots.
• ATP Bioluminescence Swabbing: Fast, but non-specific and easily confused by harmless food organic residues.

When a positive culture is finally confirmed days later, food processors are forced into catastrophic whole-batch recalls—destroying tons of wholesome food and incurring millions in financial losses.

OPTIX is an industrial-grade, 100% inline non-destructive optical screening and automated sorting system engineered for food conveyance lines moving at speeds up to $2.5\text{ m/s}$. Powered by a dual-processor architecture—an NVIDIA Jetson Orin Nano (TRINETRA Edge AI Engine) paired with an STM32 Nucleo-F446RE real-time microcontroller—OPTIX combines 11-channel strobed multimodal illumination (2x UV-A $365\text{ nm}$ + 9x VNIR), dominant tissue background suppression (Orthogonal Subspace Projection / Sparse NMF), and Conformal Prediction anchored directly to official FSSAI microbiological criteria (Tables 5A & 5B). In less than $19.2\text{ ms}$, OPTIX identifies contamination-related spectral anomalies, executes high-speed pneumatic air-jet isolation without stopping the belt, and feeds laboratory results back into the model to close the loop.

---

🥊 Head-to-Head Comparative Engineering Analysis: OPTIX vs Conventional Methods

Engineering Dimension	Conventional Lab Culture	PCR / Molecular Testing	ATP Bioluminescence Swabs	Standard RGB Machine Vision	OPTIX (Team AEROS)
Inspection Speed	Delayed $24 - 48\text{ hours}$	Delayed $4 - 8\text{ hours}$	$5 - 10\text{ minutes}$	Real-time ($30\text{ fps}$)	Real-Time Inline ($19.2\text{ ms}$ Total Latency)
Inspection Coverage	$< 0.1%$ random sample	$< 0.05%$ spot sample	$< 0.01%$ surface swab	$100%$ surface view	100% Inline Continuous Surface Screening
Sample Destructiveness	Destructive / Consumptive	Destructive / Consumptive	Swab-altered surface	Non-destructive	100% Non-Contact & Non-Destructive
Sensitivity Mechanism	Bacterial growth colonies	DNA amplification	Cellular ATP enzyme flux	Surface color / visible shape	Multimodal Hyperspectral VNIR + UV-A Autofluorescence
Sub-Surface / Invisible Anomalies	Not inline	Not inline	Blind to embedded microbes	Blind (cannot see fecal / biofilm traces)	Detects Invisible Fluorophore Emission Residuals (630–680 nm)
Food Surface Variation Robustness	N/A (Isolated in agar)	High (Purified DNA)	Very Poor (Food residue triggers false +)	Very Poor (Marbling & leaf veins cause false alarms)	Food-Specific Tissue Suppression Matrix (OSP / Sparse NMF)
Sorting Integration	Manual batch quarantine	Manual batch recall	Manual plant hold	Pneumatic / Arm diverter	Encoder-Synchronized 8-Channel Pneumatic Air-Jet Array (4.2 bar)
Uncertainty Management	Binary pass / fail	Binary pass / fail	Ambiguous RLU score	Forced binary threshold	Conformal Prediction Three-Way Triage (PASS / DIVERT / QUARANTINE)
Regulatory Benchmark	FSSAI Reference Standard	Reference Standard	Non-regulatory hygiene guide	Proprietary pixel thresholds	Direct FSSAI Regulations 2011 Tables 5A/5B Compliance Mapping
Learning Feedback Loop	None (Static manual logs)	None	None	Static model	Lab-in-the-Loop Feedback: Quarantined samples retrain edge model

---

🔬 Sensing Head Architecture & Internal Layout (Slide 2 CAD)

The OPTIX Sensing Head is a self-contained, IP65-sealed optoelectronic instrument engineered in sanitary AISI 304 Stainless Steel with standardized dimensions: $220\text{ mm}$ (Length) $\times 140\text{ mm}$ (Width) $\times 110\text{ mm}$ (Height).

       ┌────────────────────────────────────────────────────────┐
       │     OPTIX SENSING HEAD ENCLOSURE (220 × 140 × 110 mm)   │
       │                                                        │
       │   ┌────────────────┐           ┌───────────────────┐   │
       │   │  HSI / VNIR    │           │ 11-Channel LED    │   │
       │   │  Line Camera   │           │ Strobe Illuminator│   │
       │   └───────┬────────┘           └─────────▲─────────┘   │
       │           │ MIPI CSI-2                   │ Strobe PWM  │
       │           ▼                              │             │
       │   ┌────────────────┐           ┌─────────┴─────────┐   │
       │   │  Jetson Orin   │◄──UART────┤ STM32 Nucleo      │   │
       │   │  Nano (AI Core)│  Interrupt│ F446RE (Real-Time)│   │
       │   └────────────────┘           └─────────┬─────────┘   │
       │                                          │             │
       │   ┌────────────────┐                     ▼             │
       │   │ Power Mgmt     │           ┌───────────────────┐   │
       │   │ (24V→12V/5V)   │           │ MOSFET Solenoid   │   │
       │   │                │           │ Valve Drivers     │   │
       │   └────────────────┘           └─────────┬─────────┘   │
       └──────────────────────────────────────────┼─────────────┘
                                                  ▼
                                      Pneumatic Air Jets (4.2 bar)

Key Hardware Specifications:

1. Multimodal Strobe Array (11 LED Bands):
   • 2x UV-A ($365\text{ nm}$): High-power pulsed LEDs ($750\text{ mA}$) inducing autofluorescence in microbial fluorophores (NADH, riboflavin, porphyrins, and bacterial fecal residues).
   • 9x VNIR Bands ($450 - 980\text{ nm}$): Captures tissue scattering, cellular structure, hemoglobin/myoglobin, and water absorption ($950\text{ nm}$).
2. Optical Aperture Window: JGS1 optical-grade fused silica (quartz glass), broadband transmission ($250 - 2100\text{ nm} > 93%$), chemical-resistant to sanitary washdown foaming.
3. Inline Calibration Strip: Dual-region $99%$ White + $18%$ Grey reference target mounted within the line-scan field of view to perform continuous real-time flat-field calibration and flag lens fouling.
4. Pneumatic Actuation Array: 8-nozzle linear manifold ($25\text{ mm}$ pitch) driven by high-speed MOSFET solenoid switches ($18\text{ ms}$ blast at $4.2\text{ bar}$ line pressure) to divert detected items into the Quarantine Bin without disrupting the main conveyor stream.

---

🧠 System Architecture & 7-Stage End-to-End Pipeline (Slide 3)

flowchart TD
    subgraph S1 ["1. Sensing Inputs (Strobed Capture)"]
        ENC["Conveyor Optical Encoder<br/>(0.1 mm position tracking)"]
        LEDS["11 LED Strobe Bands<br/>(2 UV-A 365nm + 9 VNIR)"]
        REF["Inline Calibration Strip<br/>(99% White / 18% Grey)"]
        CAM["VNIR Line-Scan Camera<br/>(400 - 1000 nm / SWIR)"]
    end

    subgraph S2 ["2. Preprocessing (Clean & Align)"]
        DC["Dark / White Field Correction"]
        CR["Band Co-registration"]
        SQ["Strip Quality Check"]
        NF["Savitzky-Golay Denoising"]
        CUBE[("Cleaned Spatial-Spectral Cube")]
    end

    subgraph S3 ["3. Jetson Orin Nano (TRINETRA Engine)"]
        TS["Dominant Tissue Baseline Suppression"]
        UNMIX["Sparse NMF Unmixing<br/>(Bacterial, Fluorescent Residue, Wet-film)"]
        BLOB["Spatial Morphological Blobs"]
        OOD["Out-of-Distribution Check"]
    end

    subgraph S4 ["4. Decision (Risk Scoring & Triage)"]
        LGBM["LightGBM Risk Scoring Model"]
        CONF["Conformal Thresholds<br/>(τ_low = 0.35, τ_high = 0.65)"]
    end

    subgraph S5 ["5. Actuation (Belt Never Stops)"]
        STM["STM32 Nucleo-F446RE MCU"]
        VALVES["High-Speed Pneumatic Solenoids (4.2 bar)"]
        
        P_PASS["🟢 PASS: Continues to Packing"]
        P_DIV["🟠 DIVERT: High-Speed Air-Jet Reject"]
        P_QUAR["🔴 QUARANTINE: Diverted to Lab Testing"]
    end

    subgraph S6 ["6. Lab-in-the-Loop Feedback (Retrain)"]
        LAB["Laboratory Reference Testing<br/>(Plate-Count TVC / PCR)"]
        DATA["Label Validated Anomaly Data"]
        RETRAIN["Retrain Models & Recalibrate Thresholds"]
    end

    S1 --> S2
    DC --> CR --> SQ --> NF --> CUBE
    CUBE --> S3
    TS --> UNMIX --> BLOB --> OOD
    S3 --> S4
    LGBM --> CONF
    CONF -->|Risk < τ_low| P_PASS
    CONF -->|Risk > τ_high| P_DIV
    CONF -->|τ_low ≤ Risk ≤ τ_high or Strip Drift| P_QUAR
    P_DIV --> STM --> VALVES
    P_QUAR --> LAB --> DATA --> RETRAIN -.->|Deploy Updated Model| LGBM

Loading

---

📐 Mathematical Rigor & Spectral Algorithms (Slide 6)

1. Dominant Tissue Background Annihilation (Orthogonal Subspace Projection)

Clean meat muscle or spinach leaf tissue accounts for $> 95%$ of spectral reflectance. OPTIX suppresses the dominant food subspace $\mathbf{U} = \mathbf{A}{\text{food}}$ using the OSP projection operator (Harsanyi & Chang, 1994, IEEE TGRS): $$\mathbf{P}{\mathbf{U}}^{\perp} = \mathbf{I} - \mathbf{U} (\mathbf{U}^T \mathbf{U})^{-1} \mathbf{U}^T$$ Applying $\mathbf{P}{\mathbf{U}}^{\perp}$ to the observed pixel spectrum $\mathbf{y} = \mathbf{A}{\text{food}}\mathbf{x}{\text{food}} + \mathbf{A}{\text{contam}}\mathbf{x}{\text{contam}} + \mathbf{n}$: $$\mathbf{r} = \mathbf{P}{\mathbf{U}}^{\perp} \mathbf{y} = \mathbf{P}{\mathbf{U}}^{\perp} (\mathbf{A}{\text{contam}} \mathbf{x}{\text{contam}}) + \mathbf{P}{\mathbf{U}}^{\perp} \mathbf{n}$$ Because $\mathbf{P}{\mathbf{U}}^{\perp} \mathbf{A}{\text{food}} = \mathbf{0}$, dominant tissue reflectance is removed, exposing the weak microbial fluorophore residual $\mathbf{r}$.

2. Sparse NMF for Weak Signal Recovery (HyperWeak & LISTA)

To separate overlapping signatures from bacterial colonies, biofilms, and wet surface films, OPTIX implements Sparse Non-negative Matrix Factorization (IEEE doc.9834973): $$\min_{\mathbf{A}, \mathbf{X}} \frac{1}{2} |\mathbf{Y} - \mathbf{A} \mathbf{X}|F^2 + \lambda |\mathbf{X}{\text{contam}}|_1 + \beta \mathcal{R}(\mathbf{A}) \quad \text{s.t.} \quad \mathbf{A} \ge 0, ;\mathbf{X} \ge 0$$ Using LISTA (Learned Iterative Shrinkage-Thresholding Algorithm, Gregor & LeCun, 2010), the optimization is unrolled into a 4-layer feedforward network executing on Jetson Tensor Cores in $< 2.4\text{ ms}$.

3. Conformal Prediction Uncertainty Triage (Angelopoulos & Bates, 2021)

Rather than forcing a dangerous binary guess, OPTIX computes non-conformity scores and enforces quantile thresholds:

• PASS ($\hat{p} < \tau_{\text{low}} = 0.35$): Normal spectral behavior within validated bounds.
• DIVERT ($\hat{p} > \tau_{\text{high}} = 0.65$): High-confidence microbial/fecal anomaly.
• QUARANTINE ($\tau_{\text{low}} \le \hat{p} \le \tau_{\text{high}}$ OR calibration strip drift): System acknowledges statistical uncertainty, protecting consumer health by routing the item to confirmatory laboratory testing.

---

📑 FSSAI Regulatory Framework & Digital Food Safety Inspection Report (Slide 5)

OPTIX anchors its thresholds directly to the Food Safety and Standards Authority of India (FSSAI) Food Safety and Standards (Contaminants, Toxins and Residues) Regulations, 2011:

Food Category	Relevant FSSAI Standard	Reference Criterion / Limit	OPTIX Inspection Mode
Poultry / Raw Meat	FSSAI Table 5A	Total Viable Count (TVC) $\le 10^6\text{ CFU/g}$ ($\log_{10}\text{TVC} \le 6.0$); Salmonella Absent in $25\text{ g}$	HSI Reflectance ($400 - 1700\text{ nm}$)
Leafy Greens (Spinach, Lettuce)	FSSAI Table 5B	Total Viable Count (TVC) $\le 10^5\text{ CFU/g}$ ($\log_{10}\text{TVC} \le 5.0$); Listeria Absent; E. coli $< 100\text{ CFU/g}$	HSI + UV-A ($365\text{ nm}$) Autofluorescence

Digital Inspection Report Output (Slide 5 Certificate):

Every batch generates an immutable digital certificate recording:

• Report ID: TRX-2026-10-02-0011
• Production Line: Line - 1 (Leafy Greens) | Sample ID: LG-PL-000245
• Conveyor Speed: $1.5\text{ m/s}$ | Detection System: VNIR + UV-A ($365\text{ nm}$)
• Key Spectral Indicators: Peak Intensity $650-680\text{ nm}$ ($0.18$ vs ref $< 0.45$), NIR Absorption $950\text{ nm}$ ($0.28$ vs ref $< 0.50$), Spectral Anomaly Score ($0.21$ vs ref $< 0.60$)
• Detection Probabilities: Salmonella spp. ($0.08$ - Not Detected), Listeria spp. ($0.05$ - Not Detected), E. coli ($0.11$ - Not Detected) $\implies$ Overall Result: 🟢 PASS
• ISO Compliance: ISO 6579-1:2017 (Salmonella detection) and ISO 16140 series (validation of alternative microbiological methods).

---

🧪 Bench Validation Metrics & Latency Budget (Slide 6)

Benchmarked on the empirical protocol of Black et al. (2025) (Journal of Food Protection, $N = 404$ commercial carcasses and leafy green samples):

  Optical Sensor Exposure & Line Capture         :  3.5 ms
  Radiometric Flat-Fielding & Band Alignment     :  2.1 ms
  TRINETRA Sparse NMF & LightGBM Inference      :  8.4 ms
  UART Telemetry Serialization to STM32          :  1.2 ms
  STM32 Position Tracking & Queue Check          :  0.0 ms (Hardware Timer)
  Pneumatic Solenoid Coil Energization & Travel  :  4.0 ms
  ──────────────────────────────────────────────────────────
  TOTAL SYSTEM CAPTURE-TO-FIRE LATENCY           : 19.2 ms

Metric	Target Specification	Measured Bench Result	Verification Status
Goodness-of-Fit ($R^2$ on $\log_{10}\text{TVC}$)	$\ge 0.70$	$0.733$	✅ PASSED
Root Mean Square Error ($\text{RMSE}$ on $\log_{10}\text{TVC}$)	$\le 0.85\text{ CFU/g}$	$0.763\text{ CFU/g}$	✅ PASSED
False Negative Rate (FNR @ FPR $\le 2%$)	$\le 1.0%$	$0.0%$	✅ PASSED (Uncertain cases routed to QUARANTINE)
End-to-End Latency	$\le 30.0\text{ ms}$	$19.2\text{ ms}$	✅ PASSED (Operates smoothly at $2.5\text{ m/s}$)

---

📋 Bill of Materials (BOM) & Prototype Budget

Item	Component Description	Part Number / Model	Qty	Cost (INR)
1	Edge AI Compute Module	NVIDIA Jetson Orin Nano (8GB)	1	₹44,500
2	Real-Time Sync MCU	STMicroelectronics Nucleo-F446RE	1	₹1,950
3	Multispectral / VNIR Sensor	Lucid Vision / Allied Vision VNIR	1	₹68,000
4	Conveyor Optical Shaft Encoder	Omron E6B2-CWZ6C (2048 PPR)	1	₹3,400
5	UV-A High-Power Strobe LEDs	Nichia NCSU033B ($365\text{ nm}$)	2	₹1,600
6	Narrowband VNIR LEDs	Osram Oslon SSL (9 Bands)	9	₹2,700
7	Strobe MOSFET Driver Board	Custom TI TPS2814 Circuit	1	₹1,850
8	Solenoid Valve MOSFET Driver	8-Channel IRLZ44N Array	1	₹1,200
9	Industrial Step-Down Buck	Mean Well DDR-60G-24/12/5	1	₹2,100
10	High-Speed Pneumatic Valves	Festo / SMC 24V DC ($4.2\text{ bar}$)	4	₹9,600
11	Food-Grade SS304 Enclosure	Custom $220 \times 140 \times 110\text{ mm}$ IP65	1	₹4,800
12	Optical Quartz Glass Window	JGS1 Fused Silica ($4\text{ mm}$ AR coated)	1	₹3,200
13	Diffuse Calibration Reference	99% White / 18% Grey Strip	1	₹4,500
14	Mounting Gantry & Bracket	3030 T-Slot Aluminum Profile	1	₹3,100
TOTAL	Complete Working Prototype	—	—	₹1,48,400 (~$1,780 USD)

---

📂 Repository Layout

OPTIX/
├── docs/
│   ├── Architecture.md                           # 7-Stage End-to-End System Pipeline
│   ├── Bench_Validation_Plan.md                  # FSSAI Limits, TVC Testing & Latency Budget
│   ├── Sensing_Head_Architecture.md              # 220x140x110 mm CAD, 11-Channel Strobe & Optics
│   ├── Mathematical_Models_and_Spectral_Unmixing.md # HyperWeak Sparse NMF, OSP, LISTA & Conformal
│   └── FSSAI_Food_Safety_Inspection_Report.md    # Official Digital Inspection Certificate Standard
├── firmware/                                     # STM32 Nucleo-F446RE Real-Time Firmware
│   ├── Inc/
│   │   ├── main.h
│   │   ├── encoder_tracking.h                    # Optical Encoder Conveyor Position Tracking
│   │   ├── strobe_driver.h                       # 11-Channel LED Strobe Timing (2 UV-A + 9 VNIR)
│   │   └── solenoid_valve_control.h              # MOSFET Pneumatic Air-Jet Valve Driver
│   ├── Src/
│   │   ├── main.c
│   │   ├── encoder_tracking.c
│   │   ├── strobe_driver.c
│   │   └── solenoid_valve_control.c
│   └── Makefile
├── hardware/
│   ├── Sensing_Head_Signal_Chain.md              # Electronic Schematic & Signal Flow
│   ├── BOM_Component_List.md                     # Complete Bill of Materials with Pricing
│   └── Wiring_and_Integration_Guide.md           # Pinouts, Isolation & Wiring Diagrams
├── mechanical/
│   └── Sensing_Head_Mechanical_Spec.md           # SS304 Food-Grade Enclosure Specs (Slide 2 CAD)
├── simulation/
│   ├── spectral_unmixing_sim.py                  # Standalone Spectral Unmixing & OSP Bench
│   └── conformal_prediction_bench.py             # Standalone 404-Carcass Conformal Triage Bench
├── app.js                                        # Interactive Web Simulation Engine
├── index.html                                    # Zero-Scroll High-Precision Instrument UI
├── styles.css                                    # Instrument Bench Theme Styling
├── server.ps1                                    # Local Benchmark Server Launcher
├── start_demo.bat                                # Windows One-Click Offline Launcher
└── LICENSE                                       # MIT License (Team AEROS)

---

🚀 Quick Start Guide

1. Run the Python Numerical Simulations (Zero Dependencies)

# Verify Spectral Unmixing & Weak Signal Residual Extraction
python simulation/spectral_unmixing_sim.py

# Run Conformal Prediction 404-Carcass Cohort Validation Bench
python simulation/conformal_prediction_bench.py

2. Launch the Interactive Local Bench Simulator

Double-click start_demo.bat or run in PowerShell:

./server.ps1

Open your browser at http://localhost:8080.

3. Access the Live Online Web Simulator

Visit the hosted GitHub Pages deployment:
👉 https://sandy1472006/OPTIX/

---

Team AEROS · Smart India Hackathon 2026 · Team ID: 186870
Advancing Indigenous Inline Optical Inspection for Food Safety & Processing