Quality assessment and grading of onions are often subjective and vary across procurement centers, resulting in disputes and inconsistencies.

Description:	Expected Solution: Develop an AI-based mobile application that: • Uses image processing to assess onion quality. • Identifies damaged, rotten, sprouted, or undersized onions. • Estimates Grade A and URS percentages. • Generates a digital quality report instantly. • Reduces human bias and improves transparency.
Organization:	Ministry of Consumer Affairs, Food & Public Distribution
Department:	Department of Consumer Affairs (DoCA)
Category:	Software
Theme:	Smart Automationtd>
Youtube Link:	Not provided
Dataset Link:	Not provided
Contact info:	Not provided

1. Executive Snapshot

This research and solution architecture dossier defines an end-to-end strategy for building AgriGrade-AI (Onion Quality Assessment & Digital Grading System), an edge-AI mobile application designed to eliminate subjectivity, disputes, and operational opacity in onion procurement under the Department of Consumer Affairs (DoCA) Price Stabilisation Fund (PSF) [PS, RESEARCHED]. By placing a multi-onion sample on a standardized calibration tray with a physical scale reference marker, field procurement officers capture a single image [INFERENCE]. The on-device computer vision engine (YOLOv8/YOLOv11 instance segmentation running via TFLite/ONNX) automatically segments every onion, measures minor/major diameter down to millimeter accuracy, detects visual defects (rot, sprouting, mechanical skin damage, double bulbs), classifies each bulb into Grade A, URS (Under Relaxed Specifications), or Reject, and instantly computes batch percentage compositions [RESEARCHED, INFERENCE]. The application operates 100% offline in remote mandis, generating tamper-evident, geotagged digital inspection reports with QR code verification that synchronize seamlessly to central buffer stock dashboards when connectivity is restored [INFERENCE, ASSUMPTION].

Metadata Table

Parameter	Details
PS Number	SIH26031 [PS]
Title	Quality assessment and grading of onions are often subjective and vary across procurement centers, resulting in disputes and inconsistencies. [PS]
Organization	Ministry of Consumer Affairs, Food & Public Distribution [PS]
Department	Department of Consumer Affairs (DoCA) [PS]
Category	Software [PS]
Theme	Smart Automation [PS]
Deadline	30 September 2026 [PS]
Ideas Counter	0/500 [PS]

Fundamental System Objective: To deliver an offline-capable, computer vision-driven mobile grading platform that provides instant, objective, and auditable percentage estimations of Grade A, URS, and Reject onions from multi-bulb sample images, eliminating procurement center disputes and safeguarding central buffer stock quality [PS, INFERENCE].

2. Problem Deconstruction

The explicit problem statement requirement text from the Department of Consumer Affairs (DoCA) [PS] is deconstructed into atomic, verifiable requirements below:

Explicit Requirements

• REQ-E1 (Image Processing Quality Assessment): Uses image processing to assess onion quality. [PS]

  • Interpretation: The system must capture visual image data of onions and process spatial, color, and structural characteristics without relying solely on manual visual scoring. [INFERENCE]
  • Implementation Implication: High-resolution RGB image capture module on Android/iOS devices with automatic pre-processing, lighting adjustment, and contour segmentation. [INFERENCE]
  • Confidence Tag: [PS]

• REQ-E2 (Defect & Attribute Identification): Identifies damaged, rotten, sprouted, or undersized onions. [PS]

  • Interpretation: The vision algorithm must perform multi-class defect classification and spatial dimension measurement on individual onion instances within a batch image. [INFERENCE]
  • Implementation Implication: Multi-task instance segmentation network trained to detect:
    1. Rotten: Black mold (Aspergillus niger), bacterial soft rot, discolored wet lesions [RESEARCHED].
    2. Sprouted: Green foliage emergent from bulb crown [RESEARCHED].
    3. Damaged: Cut, bruised, crushed outer scales, or sunburnt/skinned layers [RESEARCHED].
    4. Undersized: Bulb diameter smaller than statutory minimum (<45 mm for strict Grade A, or <35 mm for relaxed URS) [RESEARCHED].
  • Confidence Tag: [PS]

• REQ-E3 (Grade A & URS Percentage Estimation): Estimates Grade A and URS percentages. [PS]

  • Interpretation: After detecting individual bulbs, the system must aggregate results across the sample to output precise weight or count percentages for Grade A and URS (Under Relaxed Specifications). [INFERENCE]
  • Implementation Implication: Statistical aggregator calculating $% \text{Grade A} = (N_{\text{Grade A}} / N_{\text{Total}}) \times 100$ and $% \text{URS} = (N_{\text{URS}} / N_{\text{Total}}) \times 100$, adjusted for estimated volumetric mass [INFERENCE].
  • Confidence Tag: [PS]

• REQ-E4 (Instant Digital Quality Report): Generates a digital quality report instantly. [PS]

  • Interpretation: The mobile app must compile sample inspection metrics, defect counts, grade percentages, timestamp, GPS position, and annotated images into an immediate, shareable report. [INFERENCE]
  • Implementation Implication: Local PDF/JSON report generation engine operating on-device in under 3 seconds without requiring active cloud network connectivity. [INFERENCE]
  • Confidence Tag: [PS]

• REQ-E5 (Bias Reduction & Transparency): Reduces human bias and improves transparency. [PS]

  • Interpretation: Inspection records must be objective, deterministic, verifiable by both farmers and inspectors, and protected against manual tampering. [INFERENCE]
  • Implementation Implication: Cryptographic hashing of inspection records, side-by-side visual bounding box proof on mobile screen, and sync to central DoCA audit logs. [INFERENCE]
  • Confidence Tag: [PS]

Derived & Inferred Requirements

• REQ-I1 (Physical Scale Calibration): Due to varying smartphone camera focal lengths and shooting distances, real-world diameter measurement ($mm$) requires a physical reference marker (e.g., ArUco code or fixed-size calibration mat) [RESEARCHED, INFERENCE].
• REQ-I2 (Offline-First Execution): Primary procurement occurs in rural APMC mandis with poor 3G/4G connectivity; inference and report creation must run 100% on-device [KNOWN, INFERENCE].
• REQ-I3 (Dual Policy Dynamic Thresholding): DoCA procurement guidelines alternate between strict Grade A (45–65 mm) and URS relaxed rules (35–70 mm, up to 40% surface staining) via periodic Office Memorandums; classification thresholds must be dynamically configurable via backend policy updates [RESEARCHED, INFERENCE].

3. Stakeholders & User Needs

Stakeholder	Goal	Current Pain	System Need
Farmer / Farmer Producer Organization (FPO)	Get fair, transparent price realization for onion lots based on objective quality without unfair rejections [RESEARCHED].	Subjective visual grading by procurement agents leads to high rejection rates (up to 30%), forcing distress sales in APMCs [RESEARCHED].	Instant visual report on mobile showing exact count, millimeter size distribution, and reason for URS/Reject classification [INFERENCE].
Procurement Officer (NAFED / NCCF Field Agent)	Rapidly evaluate incoming truck/trolley loads, record digital acceptance, and minimize disputes at procurement centers [RESEARCHED].	Manual sampling of 50-100 onions per trolley is slow, prone to accusations of bias, and causes physical congestion at mandis [KNOWN, INFERENCE].	Fast mobile camera scan taking <10 seconds per sample tray, auto-generating an unalterable digital slip [INFERENCE].
Mandi / Center Supervisor	Prevent corrupt grading practices, maintain throughput, and audit field surveyor decisions [INFERENCE].	Lack of visual evidence when farmers dispute quality grades; impossible to audit past manual sampling assessments [KNOWN].	Digital audit trail with geotagged, timestamped images and overlay bounding boxes stored locally and on cloud [INFERENCE].
DoCA Policy & Buffer Stock Administrators	Ensure buffer stock maintains high shelf-life, prevent rapid post-procurement rot in CWC godowns, and monitor regional quality trends [RESEARCHED].	High storage losses (up to 30-35% in warehouses) caused by accepting undetected rot/moisture-affected lots during procurement [RESEARCHED].	Real-time centralized dashboard showing batch quality breakdown across all procurement centers in Maharashtra, MP, and Gujarat [RESEARCHED, INFERENCE].

4. Requirements Traceability

Proposed System Feature	Justifying Requirement	Implementation Component
AR Camera Viewfinder with Framing Guide	REQ-E1 [PS], REQ-I1 [INFERENCE]	Flutter AR Camera plugin with distance check and tilt-sensor angle validation.
Multi-Onion Instance Segmentation Engine	REQ-E1 [PS], REQ-E2 [PS]	MobileNetV3-YOLOv8s-seg ONNX model optimized via TensorRT/TFLite.
Defect Classifier (Rot, Sprout, Damage)	REQ-E2 [PS]	Multi-head classification model operating on cropped instance bounding boxes.
ArUco / Calibration Mat Metric Converter	REQ-E2 [PS], REQ-I1 [INFERENCE]	OpenCV perimeter contour detector computing Pixel-per-Millimeter ($PPM$) ratio.
Grade A vs URS Percentage Aggregator	REQ-E3 [PS], REQ-I3 [INFERENCE]	Deterministic decision engine matching bulb measurements against active DoCA policy rules.
Instant Offline PDF & Receipt Generator	REQ-E4 [PS], REQ-I2 [INFERENCE]	Dart pdf rendering engine generating bilingual (English/Hindi/Marathi) summary slips.
Cryptographic Hash & Geotagging Module	REQ-E5 [PS]	SHA-256 payload signing with GPS coordinate and NTP time stamping.
Central Buffer Stock Analytics Portal	REQ-E5 [PS]	React/Next.js dashboard consuming synchronized batch inspection telemetry.

5. Constraints & Environment Realities

• Geography & Lighting Constraints: Procurement occurs in open-air mandis, covered sheds, or tractor trolleys under direct harsh sunlight, uneven shadows, or low-light dusk conditions [KNOWN, INFERENCE]. Mitigation: Require a foldable, low-cost blue/green matte calibration tray with built-in shade walls and standard color calibration blocks [ASSUMPTION].
• Network Connectivity Realities: Rural procurement yards in Nashik, Ahmednagar, or MP frequently suffer from complete network blackout or high latency 2G/3G connections [KNOWN, RESEARCHED]. Mitigation: All AI inference, image processing, database storage, and PDF report creation must execute 100% offline [INFERENCE].
• Hardware Limitations: Field agents use low-to-mid-tier Android smartphones (2GB–4GB RAM, MediaTek Helio or Snapdragon 6-series processors) [KNOWN, ASSUMPTION]. Mitigation: Quantize deep learning models to INT8 TFLite/NPU format, targeting model weights under 15 MB and memory usage under 250 MB [INFERENCE].
• Language & Digital Literacy: Farmers and field helpers prefer regional languages (Hindi, Marathi, Gujarati) with high visual feedback rather than text-heavy UI [KNOWN]. Mitigation: Icon-driven UI, voice guidance, and bilingual PDF printouts [INFERENCE].
• High Operational Throughput: Procurement centers handle hundreds of trolleys daily; an inspection cannot take longer than 2 minutes per sample lot [KNOWN, INFERENCE].

6. Research Questions

1. How can accurate physical bulb diameter ($mm$) be measured from a 2D smartphone image without expensive depth sensors (LiDAR/ToF)?
   • Investigation Strategy: Compare ArUco marker tags versus standardized 300x300mm foldable calibration trays with high-contrast borders [RESEARCHED].
2. What exact visual features distinguish URS (Under Relaxed Specifications) from Reject/Rotten onions under current DoCA memorandums?
   • Investigation Strategy: Analyze DoCA memorandums (e.g., June 2026/July 2026 directives) defining Grade A (45–65 mm, <10% staining) vs URS (35–70 mm, up to 40% surface staining/sunburn) [RESEARCHED].
3. How to handle overlapping bulbs placed closely on a sample tray?
   • Investigation Strategy: Evaluate Instance Segmentation (YOLOv8-seg/YOLOv11-seg) versus Watershed algorithm post-processing to separate touching onion contours [RESEARCHED].
4. What is the minimal sample size required to statistically represent a 2-tonne trolley load of onions?
   • Investigation Strategy: Reference IS 1787 / Agmark sampling guidelines prescribing a random sample of 50 to 100 bulbs per batch [RESEARCHED].

7. Data Requirements & Sources

Data Type	Purpose	Ideal Real-World Source	Hackathon Substitute	Access Method	Availability	Privacy Sensitivity	Confidence
RGB Onion Image Dataset	AI model training for instance segmentation & defect detection.	Field capture across Nashik & Lasalgaon APMC mandis [RESEARCHED].	Roboflow Agriculture Datasets, Kaggle Fresh/Rotten Produce Datasets, manual local collection [RESEARCHED].	REST API / Web Download	High (Public Kaggle/Roboflow)	Low (Public Produce)	[RESEARCHED]
DoCA Quality Standards & MAPP Memorandums	Grading rule definition (Grade A vs URS thresholds).	DoCA Official Office Memorandums & Press Information Bureau (PIB) releases [RESEARCHED].	Official PIB archive documents & DoCA press releases [RESEARCHED].	Web Scraping / Manual PDF Parsing	High (Public Government)	None (Public Policy)	[PS] / [RESEARCHED]
Procurement Center Geo-Locations	Geofencing and center assignment for inspections.	NAFED / NCCF Portal directory (e-Samridhi / e-Samyukti) [RESEARCHED].	Synthetic GeoJSON mapping of major Maharashtra/MP mandi coordinates [ASSUMPTION].	Static JSON Config File	Medium	Low	[INFERENCE]
Farmer Procurement Records	Linking sample inspection reports to farmer identity & trolley ID.	DoCA e-Samridhi / e-Samyukti REST API [RESEARCHED].	Mock SQLite/JSON API server simulating farmer registration lookup [ASSUMPTION].	Simulated REST API	Mock Data	High (PII)	[ASSUMPTION]

Data Categorization Breakdown

• A. Data Explicitly Mentioned: Onion quality images, defect types (damaged, rotten, sprouted, undersized), Grade A and URS percentage outputs [PS].
• B. Data Required by Inference: Camera calibration parameters, pixel-to-millimeter ratio, GPS coordinates, timestamp, inspection batch UUID, farmer registration ID [INFERENCE].
• C. Data That Must Be Collected: Field RGB images of Indian Red/White/Pink onion varieties under varied natural light [INFERENCE].
• D. Data That Can Be Synthetically Generated: Simulated trolley inspection logs, historical quality trends, mandi geolocation coordinates [ASSUMPTION].
• E. Data That Still Needs Validation: Real-world correlation factor between 2D surface defect area percentage and internal rot mass [NEEDS VALIDATION].

8. Existing Solutions & Gap Analysis

System / Solution	Description	Strengths	Weaknesses	Relevance	Gap Remaining	Source	Confidence
Industrial Optical Sorters (e.g., Ellips True-AI, TOMRA Sorters)	Multi-camera conveyor belt sorting machines used in commercial packing houses [RESEARCHED].	High throughput (tons/hr), 360-degree rotation, near 98% accuracy for rot/size sorting [RESEARCHED].	Extremely expensive ($50,000+), stationary, heavy industrial power required; unviable for open APMC mandis [RESEARCHED].	Benchmark for AI grading criteria.	Cannot be deployed at mobile procurement points or field mandis.	Commercial Vendor Docs [RESEARCHED]	[RESEARCHED]
Generic Plant Health Apps (e.g., Plantix, Agrio)	Mobile apps where farmers upload single leaf/crop photos to diagnose foliage diseases [RESEARCHED].	Easy mobile UX, offline-enabled classification models for foliage [RESEARCHED].	Single-item focus; no multi-item batch counting, no millimeter size measurement, no DoCA URS policy grading [RESEARCHED, INFERENCE].	Proves mobile AI feasibility for agriculture.	Lacks batch instance segmentation, physical scale calibration, and government report formats.	Public App Stores [RESEARCHED]	[RESEARCHED]
Manual Visual Inspection (Current DoCA Practice)	Human quality inspectors manually pull 10-20 onions from a trolley and assign visual grade [RESEARCHED].	Zero technology cost, fast immediate decision.	Extremely subjective, high inspector bias, frequent farmer disputes, undetected rot causes high storage losses [RESEARCHED].	Current baseline process.	Entirely lacks objectivity, transparency, and digital auditability.	DoCA / Press Reports [RESEARCHED]	[PS] / [RESEARCHED]

9. Candidate Solution Approaches

Approach 1: Lightweight Edge Computer Vision (YOLOv8s-seg + Physical Calibration Tray Mat) [RECOMMENDED]

• Mechanism: The user places a sample batch of 20–30 onions on a standardized $40 \times 40\text{ cm}$ high-contrast blue tray featuring printed Corner ArUco Markers. An on-device INT8-quantized YOLOv8s-seg model runs locally on the smartphone, segmenting each onion instance and detecting defects (rot, sprout, damage). OpenCV detects ArUco markers to compute exact pixel-per-millimeter scale and perspective transformation, measuring minor/major axis diameters for every onion. Rule-based logic applies active DoCA thresholds to compute Grade A, URS, and Reject percentages [RESEARCHED, INFERENCE].
• Required Data: 3,500 annotated onion images with polygon masks and defect labels [INFERENCE].
• Complexity: Medium | Cost: Low (Free software + ₹200 calibration tray) | Explainability: High (Annotated visual overlay on screen) [INFERENCE].
• Hackathon Feasibility: High (100% buildable in Flutter/Python/TFLite) [INFERENCE].
• Production Feasibility: Excellent (Runs offline on low-cost Android phones) [INFERENCE].

Approach 2: Multimodal Cloud Vision LLM (GPT-4o / Claude 3.5 Sonnet API)

• Mechanism: The mobile app captures a photo of onions on a tray and sends the raw image to a cloud LLM endpoint with a prompt asking for defect counts, size estimation, and Grade A/URS percentage estimation [INFERENCE].
• Required Data: Zero training data required (zero-shot prompt engineering) [KNOWN].
• Complexity: Low (API integration) | Cost: High ($0.01–$0.03 per inspection API call) [KNOWN, INFERENCE].
• Explainability: Low (LLM hallucinations on exact millimeter measurements and counts) [KNOWN].
• Hackathon Feasibility: High (Fast prototype) | Production Feasibility: Unviable (Fails completely offline in rural mandis, high operational cost, prone to hallucinated metrics) [KNOWN, INFERENCE].

Approach 3: Dual-Stage Hybrid Pipeline (Local Feature Detection + Cloud Verification)

• Mechanism: On-device Classical OpenCV contour extraction counts and measures sizes; images flagged with ambiguous defects are compressed and queued for cloud ML verification when network becomes available [INFERENCE].
• Required Data: 2,000 annotated images [INFERENCE].
• Complexity: High (Managing sync queues and dual decision loops) | Cost: Medium | Explainability: Medium [INFERENCE].
• Hackathon Feasibility: Medium | Production Feasibility: Good, but over-engineered for immediate on-field decision making [INFERENCE].

10. Recommended Approach & Why

Selected Approach: Approach 1 (Lightweight Edge Computer Vision: YOLOv8s-seg + Physical Calibration Mat Mat) [INFERENCE].

Justification & Rejection of Alternatives

1. Why Approach 2 was Rejected: Cloud Vision LLMs fail fundamental operational constraints: rural APMC mandis have zero or intermittent internet access [KNOWN, RESEARCHED]. Furthermore, LLMs hallucinate precise spatial dimensions ($mm$) and exact object counts, making them illegal/unusable for official financial transactions under DoCA [KNOWN, INFERENCE].
2. Why Approach 3 was Rejected: Dual-stage hybrid pipelines introduce unnecessary system complexity during a hackathon build without improving instant on-field decision capability. Procurement officers require immediate offline pass/fail decisions before the trolley departs [KNOWN, INFERENCE].
3. Why Approach 1 Wins: It combines deterministic physical measurement (ArUco pixel-to-millimeter geometric mapping) with deep learning pattern recognition (YOLOv8 instance segmentation for defects) [RESEARCHED]. It operates 100% offline, executes in <2 seconds on basic smartphones, provides visual explainability (polygon overlays on every bulb), and costs under ₹200 per procurement center for the physical tray [INFERENCE].

11. Technical Architecture

Layered Architecture Diagram

+-----------------------------------------------------------------------------------+
|                            PRESENTATION LAYER (Mobile App)                        |
|   +------------------------------------+    +---------------------------------+   |
|   |  AR Viewfinder & Level Indicator   |    | Digital Quality Report Viewer   |   |
|   |  (Flutter Custom Camera / Sensors) |    |  (Bilingual PDF / QR Generator) |   |
|   +------------------------------------+    +---------------------------------+   |
+-----------------------------------------------------------------------------------+
                                         |
                                         v
+-----------------------------------------------------------------------------------+
|                        ON-DEVICE EDGE AI & PROCESSING ENGINE                      |
|   +------------------------------------+    +---------------------------------+   |
|   | OpenCV Geometric Processor         |    | ONNX Runtime / TFLite Engine    |   |
|   | - ArUco Marker Detection           |    | - YOLOv8s-seg Model (INT8)      |   |
|   | - Perspective Correction (Homography)|  | - Instance Segmentation         |   |
|   | - Pixel-per-mm Calibration Ratio   |    | - Defect Head (Rot/Sprout/Cut)  |   |
|   +------------------------------------+    +---------------------------------+   |
|                                        |                                          |
|   +---------------------------------------------------------------------------+   |
|   | Deterministic DoCA Grading Logic Engine                                   |   |
|   | - Size Classifier (Grade A: 45-65mm | URS: 35-70mm | Reject: <35mm)         |   |
|   | - Defect Aggregator (% Rot, % Sprout, % Mechanical Damage)                |   |
|   +---------------------------------------------------------------------------+   |
+-----------------------------------------------------------------------------------+
                                         |
                                         v
+-----------------------------------------------------------------------------------+
|                          LOCAL STORAGE & SECURITY LAYER                           |
|   +------------------------------------+    +---------------------------------+   |
|   | SQLite Local DB (Offline Queue)    |    | SHA-256 Payload Signer          |   |
|   | Encrypted Blob Storage (Images)    |    | GPS & Timestamp Locker          |   |
|   +------------------------------------+    +---------------------------------+   |
+-----------------------------------------------------------------------------------+
                                         |
                             (Async Sync when Online)
                                         v
+-----------------------------------------------------------------------------------+
|                       CENTRAL CLOUD BACKEND (DoCA Gateway)                         |
|   +------------------------------------+    +---------------------------------+   |
|   | FastAPI REST Sync Gateway          |    | PostgreSQL + PostGIS DB         |   |
|   | JWT Authentication & Role Control  |    | Buffer Stock Analytics Engine   |   |
|   +------------------------------------+    +---------------------------------+   |
|                                        |                                          |
|   +---------------------------------------------------------------------------+   |
|   | Next.js Administrative Analytics Portal                                   |   |
|   | - Center-wise Quality Heatmaps | Buffer Loss Risk Forecasts                |   |
|   +---------------------------------------------------------------------------+   |
+-----------------------------------------------------------------------------------+

Detailed Data-Flow Narrative

1. Sample Placement & Capture: Procurement agent scoops a representative sample of ~30 onions from a farmer's trolley and places them on the standardized $40 \times 40\text{ cm}$ mat [INFERENCE].
2. Geometric Calibration: The mobile app captures the frame. OpenCV locates four ArUco markers on the tray corners, calculates the perspective transformation matrix, warps the image to a flat top-down grid, and establishes the exact scale factor: $\text{PPM} = \frac{\text{Pixel Distance}}{\text{Known mm Distance}}$ [RESEARCHED, INFERENCE].
3. Instance Segmentation & Measurement: The cropped inner tray area is fed into the TFLite YOLOv8s-seg model. The model predicts polygon masks for every individual onion. The minor axis of the minimum bounding ellipse is calculated in millimeters: $D_{\text{real}} = \frac{D_{\text{ellipse_pixels}}}{\text{PPM}}$ [RESEARCHED, INFERENCE].
4. Defect Classification: Each segmented onion patch is processed through defect classification heads predicting probabilities for Rot, Sprout, Mechanical Damage, or Sound/Clean [INFERENCE].
5. Policy Decision Engine: Active DoCA rule thresholds are applied per bulb:
   • Grade A: Diameter 45–65 mm, no rot, no sprout, damage <5% [RESEARCHED].
   • URS: Diameter 35–45 mm or 65–70 mm, or surface staining up to 40%, no rot, no sprout [RESEARCHED].
   • Reject: Diameter <35 mm, or presence of rot/sprouting/severe physical breakage [RESEARCHED].
6. Report & Sync: The app displays the total count, breakdown numbers, and Grade A / URS / Reject percentages on screen with annotated bounding boxes. A digitally signed, geotagged PDF report with a QR code is stored in local SQLite and queued for HTTPS push to the central DoCA cloud server [INFERENCE].

12. Technology Stack

Layer	Selected Technology	Why Selected	Alternative Considered	Why Alternative Rejected
Mobile Frontend	Flutter (Dart)	Single codebase for Android/iOS, high-performance native rendering, excellent camera/OpenCV bindings [INFERENCE].	React Native	Slower bridge performance for real-time frame processing [INFERENCE].
Edge AI Engine	ONNX Runtime Mobile / TFLite	Hardware-accelerated NPU/GPU execution on mobile devices, low memory footprint [INFERENCE].	PyTorch Mobile	Heavy runtime binary size (>80 MB vs <15 MB for TFLite) [INFERENCE].
Computer Vision	OpenCV (C++/Python/Dart)	Robust, deterministic ArUco marker detection, homography transformation, and ellipse fitting [RESEARCHED].	Native Canvas API	Lacks advanced geometric transformation and marker recognition functions [INFERENCE].
AI Model Architecture	YOLOv8s-seg / YOLOv11s-seg	State-of-the-art real-time instance segmentation, superior mAP for dense small objects [RESEARCHED].	Mask R-CNN	Extremely high inference latency on mobile CPU/NPU (>1.5s vs <150ms for YOLO) [RESEARCHED].
Local Database	SQLite + Drift (Dart)	Embedded relational storage, zero configuration, robust offline sync queue management [INFERENCE].	Hive / SharedPreferences	Lacks relational querying and ACID transaction guarantees for financial records [INFERENCE].
Backend API Gateway	FastAPI (Python 3.12)	Asynchronous high-throughput REST API, automatic OpenAPI docs, native integration with Python ML tools [INFERENCE].	Node.js (Express)	Less seamless integration with Python-based ML analytics and computer vision utilities [INFERENCE].
Cloud Database	PostgreSQL 16 + PostGIS	Enterprise spatial database for storing mandi geofences, inspection payloads, and time-series audit logs [INFERENCE].	MongoDB	Weaker spatial index performance and lack of strict relational constraints for procurement records [INFERENCE].
Admin Dashboard	Next.js 14 + Tailwind CSS	Server-side rendering, responsive UI components, rapid dashboard development [INFERENCE].	Pure React SPA	Slower initial load times for complex administrative reporting tools [INFERENCE].

13. AI/ML Justification

Component 1: Multi-Onion Instance Segmentation Engine

• Purpose: Detect, separate, and output individual polygon masks for every onion in a dense, touching sample batch [RESEARCHED].
• Model Type: YOLOv8s-seg (quantized to INT8 TFLite) [RESEARCHED].
• Input: $640 \times 640 \times 3$ warped RGB image of the calibration tray.
• Output: $N$ instance masks, bounding boxes, and confidence scores ($0.0 - 1.0$).
• Data Required: 3,500 annotated multi-onion images with polygon boundary ground truth [INFERENCE].
• Why AI: Classical watershed algorithms fail when onions have varying skin tones, overlapping dry outer scales, or shadow clutter [RESEARCHED, INFERENCE].
• Evaluation: Mask mAP@50-95 > 88%, Inference time <200ms on Snapdragon 680 [INFERENCE].
• Failure Mode: Extreme overlapping (onions stacked 2-deep). Mitigation: User interface enforces "Single Layer Placement" rule via camera frame checks [INFERENCE].
• Human Oversight: Officer can tap on screen to manually adjust or delete misaligned bounding polygons prior to final report lock [INFERENCE].

Component 2: Defect Classification Head

• Purpose: Classify surface defects on each segmented onion patch into Rot, Sprout, Mechanical Damage, or Sound [RESEARCHED].
• Model Type: Lightweight MobileNetV3-Small classifier backbone [RESEARCHED].
• Input: Bounding box crop of individual onion instance ($128 \times 128 \times 3$).
• Output: Softmax probability vector across 4 defect classes.
• Data Required: 6,000 cropped onion images labeled by agricultural quality experts [INFERENCE].
• Why AI: Visual patterns of black mold (Aspergillus), green sprout shoots, and skin breaks are complex high-level texture variations impossible to capture with fixed color thresholds [RESEARCHED, INFERENCE].
• Evaluation: Precision > 92%, Recall > 90% on Rot/Sprout classes [INFERENCE].
• Failure Mode: Internal rot with zero outer skin discoloration. Mitigation: Explicit report disclaimer stating visual inspection covers surface characteristics only [INFERENCE].
• Human Oversight: Any high-uncertainty classification (confidence < 60%) prompts manual inspector confirmation [INFERENCE].

Where AI Should NOT Be Used

• Final Diameter Measurement ($mm$): Calculated purely through OpenCV spatial ellipse fitting and deterministic ArUco geometric scaling ratio [RESEARCHED, INFERENCE].
• Grade A vs URS Percentage Logic: Calculated using deterministic mathematical formulas matching measured diameters and defect flags against active DoCA policy parameters [INFERENCE].
• PDF Slip Cryptographic Hashing: Built using standard SHA-256 cryptographic libraries [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

Entity Relationship Schema (Local SQLite & Cloud PostgreSQL)

[ProcurementCenter] 1 --- * [FieldOfficer]
[FieldOfficer]        1 --- * [InspectionBatch]
[InspectionBatch]     1 --- * [OnionInstance]
[InspectionBatch]     1 --- 1 [QualityReport]

• InspectionBatch Table: batch_uuid (TEXT PK), officer_id (TEXT), mandi_id (TEXT), farmer_id (TEXT), trolley_number (TEXT), timestamp_utc (INTEGER), latitude (REAL), longitude (REAL), total_count (INTEGER), grade_a_count (INTEGER), urs_count (INTEGER), reject_count (INTEGER), grade_a_percent (REAL), urs_percent (REAL), doca_policy_version (TEXT), sync_status (INTEGER: 0=Pending, 1=Synced), payload_hash (TEXT).
• OnionInstance Table: instance_id (TEXT PK), batch_uuid (FK), major_diameter_mm (REAL), minor_diameter_mm (REAL), defect_type (TEXT: Sound/Rot/Sprout/Damage), defect_confidence (REAL), assigned_grade (TEXT: Grade_A/URS/Reject), crop_image_path (TEXT).

Offline Synchronization & Sync Strategy

1. Local Autonomy: Inspections are executed entirely against local SQLite. Raw cropped images and JSON payloads are saved to device storage [INFERENCE].
2. Sync Queue: When an inspection completes, a record is added to sync_queue with status PENDING.
3. Background Worker: Flutter WorkManager triggers periodic sync checks every 15 minutes. When Wi-Fi/4G becomes available, it executes an idempotent HTTP POST to /api/v1/sync/batch.
4. Bandwidth Reduction: Only the compressed summary JSON and thumbnail image of the annotated tray (<200 KB) are pushed during cellular sync; full-resolution crops remain local unless explicitly requested during an audit [INFERENCE].
5. Conflict Handling: Server utilizes Immutable Write-Once semantics (batch_uuid primary key). Re-sent batches with identical UUIDs are safely ignored (Idempotency) [INFERENCE].

15. Security, Privacy & Governance

• Personal Data Protection: Farmer identifiers (Aadhaar/e-Samridhi ID, Phone Number) are encrypted at rest using AES-256 in local SQLite and hashed in audit logs [INFERENCE].
• Anti-Tampering & Auditability: Every inspection payload is signed with SHA-256 combining batch_uuid + timestamp + GPS_coords + counts. Any manual edit of local SQLite tables invalidates the signature, flagging the record during cloud sync [INFERENCE].
• Geofencing Verification: The mobile app cross-references device GPS coordinates with designated APMC mandi polygon boundaries to prevent off-site fake inspections [INFERENCE].
• Regulatory Compliance & Prototype Disclaimer: All generated digital slips include an explicit disclaimer: "This digital inspection report is an AI-assisted estimate generated for SIH demonstration/prototype evaluation under DoCA guidelines. Official financial settlement remains subject to authorized procurement officer sign-off." [INFERENCE].

16. MVP Scope (Hackathon Build)

In Scope (MUST BUILD & SHOULD BUILD)

• Mobile App (Android APK): Built in Flutter with camera capture view and real-time tray overlay guide [INFERENCE].
• On-Device Computer Vision Engine: OpenCV ArUco calibration mat detector ($PPM$ metric calculator) + INT8 TFLite YOLOv8s-seg model [RESEARCHED, INFERENCE].
• Defect & Size Detector: Auto-segmentation of up to 30 onions per frame, outputting minor diameter in $mm$, detecting Rot, Sprout, Undersized [RESEARCHED, INFERENCE].
• DoCA Policy Logic Engine: Dynamic classification into Grade A, URS, and Reject based on 35–70 mm diameter and defect rules [RESEARCHED, INFERENCE].
• Instant Digital Quality Report: Local PDF receipt generation containing visual annotated sample image, grade percentages, timestamp, GPS, and QR code [INFERENCE].
• Offline Storage & Sync Engine: Local SQLite storage with simulated background REST API sync [INFERENCE].
• Central Analytics Portal: Web dashboard (Next.js) displaying mandi-wise quality metrics, Grade A vs URS ratio, and batch audit logs [INFERENCE].

Described Only / Out of Scope (NICE TO HAVE & DO NOT BUILD)

• Internal Rot / Hyperspectral Analysis: Requires NIR/X-ray sensors unavailable on standard smartphones [RESEARCHED].
• Direct Bank Account Payment Gateway (e-RUPI / PFMS Integration): Out of hackathon scope; simulated via mock e-Samridhi API payloads [ASSUMPTION].
• Automated Conveyor Motor Hardware Integration: Software-only focus per PS category [PS].

17. Live Demo Simulation Plan

To present a flawless, credible live demonstration before SIH judges without requiring live mandi connectivity or physical agricultural field setups:

1. Physical Demonstration Rig:
   • The Tray: Bring a physical $40 \times 40\text{ cm}$ foldable blue calibration tray with printed ArUco corner markers [ASSUMPTION].
   • The Test Onions: Prepare 12–15 real physical onions representing diverse test cases:
     • 3x Large Grade A onions (55–60 mm diameter, clean skin) [RESEARCHED].
     • 3x URS Onions (38–42 mm diameter, minor surface staining) [RESEARCHED].
     • 3x Undersized Rejects (<30 mm diameter) [RESEARCHED].
     • 3x Defective Rejects (1 sprouted onion with visible green shoot, 1 discolored rotten onion, 1 cut/damaged bulb) [RESEARCHED].
2. Live Inspection Walkthrough:
   • Open AgriGrade-AI app on Android smartphone. Point camera at the tray under ambient hall light.
   • Show AR camera feedback: "Tray Detected, Distance OK, Tilt Angle OK".
   • Tap "Execute Instant Quality Assessment".
   • Show processing speed (<2 seconds on-device).
   • Display interactive output screen: show every onion outlined in colored polygons (Green = Grade A, Yellow = URS, Red = Reject).
   • Tap on individual onion mask to show live measurements: e.g., "Bulb #4: Minor Axis 52.4 mm, Defect: None -> Grade A", "Bulb #9: Minor Axis 31.2 mm, Defect: Undersized -> Reject".
   • Show aggregated summary: Grade A: 40%, URS: 35%, Reject: 25%.
3. Offline & Sync Simulation:
   • Turn phone onto Airplane Mode. Perform a second inspection. Show that PDF report generates instantly offline.
   • Turn Wi-Fi back ON. Show live background sync pushing payload to the web analytics dashboard running on a laptop screen [INFERENCE].

18. Implementation Roadmap

Phase 1: Data Collection, Annotation & Model Training (Target: June 30, 2026)

• Collect 2,500 multi-onion images under diverse lighting conditions [INFERENCE].
• Annotate polygon masks and defect labels in Roboflow [RESEARCHED].
• Train YOLOv8s-seg instance segmentation and MobileNetV3 defect classifier in PyTorch [RESEARCHED].
• Export and quantize models to INT8 TFLite format; target model size <15 MB [INFERENCE].

Phase 2: Mobile App Development & CV Integration (Target: July 31, 2026)

• Develop Flutter UI for camera capture, AR guidance, and inspection result screens [INFERENCE].
• Integrate OpenCV C++/Dart plugin for ArUco marker detection and homography warping [RESEARCHED].
• Implement TFLite inference pipeline and deterministic DoCA policy decision engine [INFERENCE].
• Build local SQLite database, PDF report generator, and cryptographic hash locker [INFERENCE].

Phase 3: Cloud Backend & Dashboard Development (Target: August 31, 2026)

• Build FastAPI backend for batch sync, JWT auth, and PostgreSQL storage [INFERENCE].
• Develop Next.js administrative dashboard with interactive spatial maps and mandi analytics [INFERENCE].
• Implement background sync queue and offline-first conflict resolution [INFERENCE].

Phase 4: Field Testing, Optimization & Demo Finalization (Target: September 30, 2026)

• Conduct field validation tests under varied natural sunlight and overlapping scenarios [INFERENCE].
• Optimize mobile NPU execution time to <1.5 seconds per frame [INFERENCE].
• Finalize SIH live demo physical rig, bilingual printouts, and judge presentation deck [INFERENCE].

19. Testing & Validation Plan

• Functional Testing: Verify that ArUco metric scaling accurately measures physical objects of known size (e.g., a $50.0\text{ mm}$ test disk) within $\pm 1.0\text{ mm}$ error margin [RESEARCHED, INFERENCE].
• AI Model Metrics:
  • Instance Segmentation Mask mAP@50 > 90% on test dataset [INFERENCE].
  • Defect Classification Precision > 92% for Rot and Sprout categories [INFERENCE].
• False Positive / False Negative Testing:
  • False Positive Rot Test: Verify clean purple/red onion skin reflections are not misclassified as rot.
  • False Negative Size Test: Verify onions measuring $34.8\text{ mm}$ are correctly flagged as Undersized Reject and not miscategorized as URS Grade ($35.0\text{ mm}$ boundary) [RESEARCHED, INFERENCE].
• Performance Testing:
  • End-to-end local inference + PDF creation time < 3.0 seconds on low-tier Android device (Snapdragon 680, 3GB RAM) [INFERENCE].
  • Mobile app RAM consumption < 250 MB during camera execution [INFERENCE].
• Offline Resilience Testing: Perform 50 sequential offline inspections in Airplane Mode; verify zero data loss upon reconnecting and executing server sync [INFERENCE].

20. Risk Register

Risk	Likelihood	Impact	Severity	Mitigation Strategy
Harsh Natural Lighting & Shadows in Mandis	High	High	High	Foldable calibration tray includes 3-sided matte shade walls; software includes auto-exposure normalization [INFERENCE].
Overlapping / Staked Onions on Tray	Medium	High	High	Mobile camera viewfinder displays real-time warning ("Spread Onions in Single Layer") if bounding box overlap ratio > 15% [INFERENCE].
Hardware Performance Bottlenecks on Low-End Phones	Medium	Medium	Medium	Quantize neural networks to INT8 TFLite, utilize NPU hardware acceleration, resize camera input frame to 640x640 [INFERENCE].
Changing DoCA Procurement Policy Thresholds	High	Medium	Medium	Store grading rules (size limits, allowed defect %) in a dynamic local JSON configuration synced from server [RESEARCHED, INFERENCE].
Resistance / Distrust from Field Procurement Agents	Medium	High	High	Provide visual bounding box overlay proof on mobile screen and allow manual override/audit mode with inspector signature [INFERENCE].

21. Cost & Scalability

Hackathon Prototype Layer

• Development Hardware: Existing student laptops & Android mobile phones (₹0) [INFERENCE].
• Calibration Trays: 2x Custom-printed foldable plastic trays with ArUco markers (~₹400) [INFERENCE].
• Cloud Infrastructure: Free-tier Render/Vercel deployment + Supabase PostgreSQL free tier (₹0) [INFERENCE].
• Total Hackathon Cost: ~₹400 ($5 USD) [INFERENCE].

Production Deployment Layer (Per 1,000 Procurement Centers)

• Mobile Hardware: Ruggedized Android Handhelds (if field agents do not use BYOD): ₹10,000 per center [INFERENCE].
• Calibration Equipment: Heavy-duty matte PVC calibration tray: ₹500 per center [INFERENCE].
• Cloud & Server Infrastructure: Managed AWS/GCP Kubernetes Cluster + PostgreSQL RDS + S3 Blob Storage (~₹1,500,000 / $18,000 USD annually across all centers) [INFERENCE].
• Scalability Bottleneck: Network bandwidth for cloud sync is minimized by pushing only lightweight JSON metadata (~15 KB/batch) during operational hours, deferring high-res image uploads to night Wi-Fi windows [INFERENCE].

22. Differentiation

Parameter	Generic Hackathon Submissions	Proposed AgriGrade-AI Solution
Metric Scale Measurement	Guesses fruit size using uncalibrated pixel bounding boxes (fails when camera distance changes) [INFERENCE].	ArUco Homography Calibration Tray: Computes exact millimeter minor/major axis diameters invariant to camera distance/tilt [RESEARCHED, INFERENCE].
Domain Policy Alignment	Applies generic "Good vs Bad" fruit labels [INFERENCE].	DoCA Policy Integration: Directly implements official Grade A vs URS (Under Relaxed Specifications) multi-tier rules per official memorandums [RESEARCHED, INFERENCE].
Batch Capability	Analyzes a single onion placed in isolation [INFERENCE].	Multi-Onion Instance Segmentation: Evaluates 25–30 onions simultaneously in one camera capture [RESEARCHED, INFERENCE].
Network Dependency	Requires active cloud API access (fails in mandis) [KNOWN, INFERENCE].	100% Edge-AI Offline Execution: Model inference, PDF report generation, and cryptographic locking run completely offline [INFERENCE].
Transparency & Audit	Displays non-verifiable on-screen text [INFERENCE].	Cryptographically Signed PDF Receipts: Includes GPS location, timestamp, QR verification, and visual annotated proof [INFERENCE].

23. Future Scope

1. Hyperspectral / Near-Infrared (NIR) Mobile Accessories: Integrating low-cost clip-on NIR optical sensors to detect internal rot (Fusarium basal rot / internal neck rot) invisible to standard RGB surface cameras [RESEARCHED].
2. Automated Mandi Conveyor Line Integration: Adapting the edge computer vision model to high-speed Raspberry Pi / Jetson Orin industrial cameras mounted above APMC sorting chutes [RESEARCHED, INFERENCE].
3. Predictive Storage Shelf-Life Modeling: Combining batch defect percentages with ambient warehouse temperature/humidity sensors to forecast maximum safe storage days before rot outbreak in CWC buffer godowns [RESEARCHED, INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions Log

• [ASSUMPTION-1] Field procurement officers possess Android smartphones running Android 9.0+ with at least 2GB RAM and an 8MP camera.
• [ASSUMPTION-2] A 30-onion random sample drawn from a tractor trolley provides a statistically representative quality estimation for the entire lot under IS 1787 standards.
• [ASSUMPTION-3] Field mandis will adopt a standardized ₹200 foldable calibration tray mat for batch inspections.
• [ASSUMPTION-4] Farmer ID lookup and payment reconciliation APIs are simulated via mock JSON endpoints mimicking e-Samridhi / e-Samyukti architecture.

Consolidated Validation Needs Log

• [NEEDS VALIDATION-1] Exact correlation factor between 2D visible surface rot percentage and total volumetric decay mass in Indian Red Onions under real warehouse conditions.
• [NEEDS VALIDATION-2] Specific official maximum permissible error tolerances allowed by DoCA for automated mobile quality estimation in official procurement disputes.

Final Recommendation

The team should proceed with building AgriGrade-AI using Approach 1 (Edge-AI Mobile App + ArUco Calibration Mat). This approach directly solves the core problem of subjective grading, complies with official DoCA Grade A vs URS specifications, operates entirely offline in remote APMC mandis, and delivers a highly visual, auditable demo for SIH judges [PS, RESEARCHED, INFERENCE].

Recommended Stack

• Mobile App: Flutter (Dart) [INFERENCE]
• Edge ML Engine: TFLite / ONNX Runtime Mobile [INFERENCE]
• Computer Vision: OpenCV (ArUco, Perspective Warp, Ellipse Fitting) [RESEARCHED]
• AI Architecture: YOLOv8s-seg (INT8 Quantized) + MobileNetV3 Classifier [RESEARCHED]
• Local Database: SQLite + Drift [INFERENCE]
• Cloud Backend: FastAPI + PostgreSQL / PostGIS [INFERENCE]
• Admin Portal: Next.js 14 + Tailwind CSS [INFERENCE]

Biggest Dependency

Availability of a diverse, well-annotated dataset of Indian onion varieties (Nashik Red, Light Red, White) featuring varying degrees of rot, sprouting, mechanical skin damage, and size distributions under natural outdoor lighting [RESEARCHED, INFERENCE].

Biggest Risk

Harsh outdoor sunlight, deep shadows, and severe onion overlapping on the sample tray causing segmentation boundaries to merge or misclassify surface discolouration as rot [INFERENCE].

Biggest Technical Challenge

Achieving real-time, high-precision instance segmentation and metric diameter calculation ($mm$) for 30 touching onions on low-end mobile hardware in under 2 seconds while running 100% offline [RESEARCHED, INFERENCE].

First 5 Actions

1. Source & Annotate Dataset: Download public produce datasets (Roboflow/Kaggle) and collect 500 local onion images; annotate polygon masks for onions, rot, and sprouts [INFERENCE].
2. Fabricate Calibration Mat: Print a prototype $40 \times 40\text{ cm}$ blue calibration mat with four 50mm ArUco corner markers [RESEARCHED].
3. Build OpenCV Calibration Pipeline: Develop Python/Dart script to detect ArUco markers, apply homography perspective correction, and verify $PPM$ distance measurement accuracy on physical test objects [RESEARCHED, INFERENCE].
4. Train & Quantize YOLOv8s-seg: Train YOLOv8s-seg on annotated onion dataset, evaluate mAP, and export to INT8 quantized TFLite format [RESEARCHED].
5. Develop Flutter Offline Core: Construct the Flutter camera UI, TFLite integration, SQLite schema, and local PDF receipt generator [INFERENCE].

Validation Before Coding

• Verify that OpenCV ArUco detection reliably finds corner markers on the printed tray under direct sunlight and high shadow conditions with <2% dimensional measurement error [RESEARCHED, INFERENCE].

Buildability Score /100

88/100 (Highly buildable by a 4-6 student team in the available timeframe using open-source Flutter, YOLOv8, OpenCV, and FastAPI technologies) [INFERENCE].

Confidence

High Confidence (89%) in technical feasibility, domain accuracy, and alignment with DoCA requirements [PS, RESEARCHED, INFERENCE].

Research Sources

1. Hindustan Times (July 31, 2026): "Onion procurement price hiked 6th time this season"
   URL: https://www.hindustantimes.com/india-news/onion-procurement-price-hiked-6th-time-this-season-101785438050757.html [1.1]
   Source Type: Reputable Secondary News Source
   Supported Claim: Supports DoCA MAPP pricing updates, discontinuation/re-introduction of Grade-URS onions, and buffer stock targets.
   Reliability: High

2. Hindustan Times (June 4, 2026): "Centre relaxes quality norms for onion procurement from Nashik farmers"
   URL: https://www.hindustantimes.com/cities/pune-news/centre-relaxes-quality-norms-for-onion-procurement-from-nashik-farmers-101780515349362.html [1.4]
   Source Type: Reputable Secondary News Source
   Supported Claim: Supports exact Grade A size definitions (45-65mm vs relaxed 35-70mm) and URS specifications (up to 40% staining, 10% sunburn).
   Reliability: High

3. Times of India (July 30, 2026): "Centre hikes onion procurement rate to Rs 23.35/kg, but tightens quality norms"
   URL: https://timesofindia.indiatimes.com/city/nashik/centre-hikes-onion-procurement-rate-to-rs-23-35/kg-but-tightens-quality-norms-farmers-fume-at-meagre-increase/articleshow/132747274.cms [2.1]
   Source Type: Reputable Secondary News Source
   Supported Claim: Confirms NAFED and NCCF operational procurement centers in Nashik district and farmer dissatisfaction with visual grading opacity.
   Reliability: High

4. Press Information Bureau (PIB) India (July 4, 2026): "Government raises onion procurement price by 13% to ₹2125 per quintal..."
   URL: https://www.pib.gov.in/PressReleasePage.aspx?PRID=2280999&reg=48&lang=1 [2.4]
   Source Type: Official Government Portal
   Supported Claim: Confirms Department of Consumer Affairs Price Stabilisation Buffer stock operations and national production figures.
   Reliability: Official / Highest

5. Ultralytics YOLO Documentation (2026): "YOLO Computer Vision Tasks Overview & Model Specs"
   URL: https://docs.ultralytics.com/tasks [1.3]
   Source Type: Official Technology Documentation
   Supported Claim: Supports real-time instance segmentation (YOLOv8s-seg/YOLOv11s-seg) performance and mobile edge deployment capability.
   Reliability: Official / Technical

6. Roboflow Computer Vision Blog (December 2023): "Monitoring Plant Growth using Computer Vision and Reference Markers"
   URL: https://blog.roboflow.com/monitor-plant-growth/ [3.2]
   Source Type: Industry Technical Publication
   Supported Claim: Supports camera distance/scale calibration using fixed-size reference markers to compute pixel-to-millimeter physical conversion.
   Reliability: High Technical