Batangas State University · College of Engineering BS Computer Engineering · 2026

Enhancing steel
quality control through
computer vision and machine learning.

A multi-sensor defect detection system combining a USB webcam and thermal infrared camera, deployed as a web-based quality management platform.

Eugene Gabriel NarvadezSystem architect
Jazther HebradoHardware integration
Ana Marie RamosDataset & ML
Lynzy Anne RevillaCompliance & QA
Thesis Adviser
Dr. Jen Aldwayne Delmo
Scroll to begin
CHAPTER 01 01

Introduction

Every industry runs on steel. Every car frame, ship hull, bridge member, and aircraft component begins as a sheet on a rolling line — and every one of them fails when a surface defect goes uncaught.

We built a system that catches them — multi-sensor, ISO-aligned, and running on consumer-grade hardware for under ₱31,000.

Background Significance Objectives Scopes & Limitations
Background

Steel underpins safety-critical industries.

Automotive frames, ship hulls, bridge members, aircraft components — defective steel propagates downstream into systems where failure has real consequences. Surface defects compromise corrosion resistance, introduce stress concentrations, and weaken load-bearing parts.

Scratches

Linear abrasions from foreign objects or interlayer friction during coiling — disrupt protective oxide layers and accelerate corrosion in service.

Dents

Localized depressions from impact pressure — introduce stress concentrations that weaken load-bearing parts under fatigue loading.

Background · Manual inspection
Why manual inspection is not effective

Three failure modes the literature consistently identifies.

Studies have confirmed that manual visual examination is subject to subjective judgment, generally lacks precision for small or subtle defects, and cannot match the pace of modern continuous-rolling lines (Wang et al., 2021; Zheng et al., 2021).

FATIGUE

Performance decays within minutes

Even well-trained inspectors miss significant fractions of real defects after sustained visual work; degradation begins within minutes of continuous inspection.

SUBJECTIVITY

Inspectors disagree on borderline cases

Manual classification depends on training, experience, and individual judgment — producing inconsistent results across shifts and operators.

SPEED

Cannot match production-line pace

Human perception has hard physical limits; high-speed rolling lines produce sheets faster than any operator can verify reliably.

Background · Prior approaches
What past researchers have tried

Four categories — and a shift to learning since 2021.

Tang et al. (2023) categorize automated steel-surface inspection methods into four broad approaches. The most notable trend across recent literature is the rapid rise of learning-based methods, particularly from 2021 onward.

1

Statistical

Thresholding, histogram-based filters. Fragile to lighting variation and surface noise.

2

Spectral

Fourier and wavelet-based filtering. Sensitive to scale and orientation.

3

Texture segmentation

GLCM, gray-level run-length. Works in controlled conditions, struggles with overlapping defects.

4

Machine learning

CNNs, YOLO variants, transformers. Significant performance gains since 2021 — the trajectory this thesis builds on.

Significance
Why this work matters

Beneficiaries across four constituencies.

A working multi-sensor defect detection system on consumer-grade hardware has measurable impact at the production line, downstream in end-user industries, in the research community, and for the broader workforce and environment.

INDUSTRY

Steel manufacturing

Replaces subjective manual inspection with a data-driven pipeline. Reduces inspection downtime, material waste, and misclassification costs — improving competitiveness.

DOWNSTREAM

End-user industries

Construction, transportation, and infrastructure benefit from fewer defective sheets reaching their lines — mitigating structural-failure risk that carries public-safety consequences.

RESEARCH

Research community

Contributes a replicable multi-sensor fusion architecture, an open YOLO benchmark on a custom steel dataset, and an ISO-aligned web platform pattern for future systems.

SOCIETY

Society & environment

Reduces worker exposure to hazardous inspection environments. Earlier defect catch means less downstream rework and lower material waste — a more sustainable production cycle.

Objectives
What we set out to do

Five deliverables. One integrated system.

Dataset
To acquire the dataset necessary for training the models, and to verify its results compared to industry standards.
Detection
To determine various defects in steel sheets through deep learning models and optimal camera placement.
Web application
To develop a web application that can be integrated with manufacturing systems.
Prototype
To develop a prototype that demonstrates the end-to-end defect detection pipeline in a controlled environment.
Compliance
To evaluate the system against ISO 9001 and ISO 10012, ensuring classification accuracy meets a minimum threshold acceptable in steel sheet manufacturing.
Scopes & Limitations
Bounded by

What's in. What's deliberately out.

Every claim in this thesis is bounded by the scope established at the outset. The system was designed and evaluated within these boundaries — not beyond them.

IN SCOPE
  • Surface defects only — scratches and dents
  • Detectable defect size: 5 mm to 50 mm
  • Conveyor speed: 12 cm/s in controlled lab
  • USB webcam + MLX90640 thermal IR sensors
  • YOLOv8 / YOLOv11 / YOLO26 comparative training
  • Web app integration with manufacturing workflow
  • ISO 9001:2015 + ISO 10012:2003 compliance review
OUT OF SCOPE
  • Pitting, rolled-in scale, laminar cracks, oxide streaks
  • Internal or subsurface material analysis
  • Real-time factory integration with PLC / SCADA
  • Industrial-grade hardware design and certification
  • Lighting calibration across multiple inspection stations
  • Thermal sensor radiometric calibration
  • Heat, dust, smoke, vibration — environmental robustness
CHAPTER 02 02

Methodology

How the system was designed, built, and evaluated.

Research Design Design Constraints & Tradeoffs Conceptual Framework Technical Design & Procedure Data Collection Testing & Evaluation Procedure Statistical Treatment
Research Design
How we approached the work

Applied + experimental research with quantitative evaluation.

A design-and-build study that integrates computer vision and thermal sensing into a working prototype, then evaluates it against industry benchmarks. The work is comparative (three YOLO variants), empirical (held-out test set), and standards-grounded (ISO 9001 + ISO 10012).

APPROACH

Multi-sensor fusion

USB webcam + MLX90640 thermal IR combined through a soft-voting ensemble. Each modality calibrated against per-sensor validation accuracy.

COMPARISON

Three YOLO variants benchmarked

YOLOv8, YOLOv11, and YOLO26 trained under identical conditions on the same dataset. Best-performing variant selected for deployment.

EVIDENCE

Quantitative metrics

Precision, Recall, F1, mAP@0.5, mAP@0.5:0.95, FDR, MDR — all measured on a strictly held-out test set.

STANDARDS

ISO-aligned review

Twelve clauses across ISO 9001:2015 and ISO 10012:2003 reviewed with explicit Compliant / Partial / Not-compliant findings.

Design Constraints & Tradeoffs
Three architectures evaluated

Design 3 was chosen — Multi-Sensor Fusion.

We considered three candidate system architectures and scored each against five weighted criteria: classification accuracy (30%), robustness to sensor noise (25%), processing speed (20%), implementation complexity (15%), and dataset suitability (10%). Design 3 took the lead on the two heaviest-weighted dimensions.

Design 1
Architecture
Linear IPO
Weighted score
6.35

Single linear pipeline — sensors feed directly into the detection pipeline and out to a verdict. Simplest to implement but has no error correction or refinement.

Design 2
Architecture
Sequential filtering
Weighted score
6.90

Hierarchical inspection — webcam first, thermal second. Early rejection saves cycles but no probabilistic resolution of borderline cases.

Design 3
Architecture
Multi-sensor fusion
Weighted score
7.05

Two sensors feed a soft-voting ensemble. Highest scores on accuracy and noise robustness — the two criteria that together carry 55% of the weight.

WHY DESIGN 3

It wins where the score weighs heaviest.

Classification accuracy and robustness to sensor noise together account for 55% of the total weight — the two dimensions where missing the call costs the most in production. Design 3 scored highest on both, even at the cost of higher implementation complexity and slower per-cycle throughput. These tradeoffs are manageable: a standard PC platform handles the simultaneous sensor evaluation, and the controlled laboratory environment supports the precise sensor alignment the ensemble requires.

Conceptual Framework
What we built

Steel IRIS — five-step inspection pipeline.

USB webcam runs YOLOv11. MLX90640 catches subsurface anomalies. Soft-voting ensemble fuses them. Steel IRIS records every verdict with an immutable evidence trail.

STEP 1
Sheet entry
Loaded onto prototype conveyor at 12 cm/s, triggers inspection cycle.
STEP 2
Relay capture
USB webcam (Station 1) captures first; MLX90640 thermal (Station 2) captures as the sheet advances.
STEP 3
Per-modality inference
YOLOv11 detects bounding boxes; thermal evaluator flags anomalies.
STEP 4
Soft-voting fusion
Weighted ensemble (0.65 visual, 0.35 thermal) yields a fused probability.
STEP 5
Verdict + record
Annotated record streamed to Steel IRIS, kanban workflow, immutable scan log.
2.4 · Model selection
Three YOLO variants benchmarked

YOLOv11 won. Same data, same hyperparameters.

YOLOv8
mAP@0.5
0.791
F1 score
0.776
YOLOv11
mAP@0.5
0.904
F1 score
0.889
YOLO26
mAP@0.5
0.847
F1 score
0.835
2.5 · Visual detection
YOLOv11 in action

Bounding-box detection across a 5–50 mm defect range.

Per-class F1: Dent 0.877 / Scratch 0.902. Inference 9.3 ms per image on the mini-PC.

scratch · 0.91
dent · 0.84
scratch · 0.88
2.6 · Thermal sensing
MLX90640 · 32×24 array

Thermal catches what vision can't.

A subsurface dent may not show a visible boundary under overhead LED, but it shifts local heat-conduction. Per-pixel EMA baseline + 2.5σ + 1.0°C + 3-frame persistence flags real anomalies, not sensor noise.

α=0.02
EMA smoothing
≥3
Frames of persistence
MLX90640 · 32×24 · 8 Hz
Data Collection
Custom-collected + supplemented

500 images. 682 bounding boxes. Two classes.

Built on the prototype rig under controlled LED illumination. Supplemented with a remapped NEU-DET subset to expand defect-morphology coverage. Roboflow handled bounding-box annotation, preprocessing, and augmentation — horizontal/vertical flips, rotational variation, contrast adjustment, and synthetic noise.

0
Total annotated images
0
Bounding-box annotations
350/100/50
Train / Validation / Test split
0
Defect classes (dent, scratch)
Testing & Evaluation Procedure
How we measured success

Our Process.

Evaluation runs in two stages: model-level (each YOLO variant alone), then system-level (the integrated multi-sensor pipeline). The test partition is maintained strictly separate from training and validation throughout development.

STAGE 1
Model evaluation
Each YOLO variant scored on held-out test set. Confusion matrices computed per variant.
STAGE 2
Model selection
YOLOv11 chosen on superior mAP@0.5 and F1. Integrated into fusion pipeline.
STAGE 3
Per-sensor accuracy
Webcam-only and thermal-only baselines measured against ground truth on the same test sheets.
STAGE 4
System acceptance
Fused pipeline evaluated across all four dimensions: accuracy, FDR, MDR, processing time per image.
STAGE 5
Compliance review
12 ISO 9001 / 10012 clauses scored Compliant / Partial / Not-compliant with explicit evidence.
TEST SET

50 images held out

42 No-Good · 8 Good. Strictly separated from training and validation throughout all development activities.

GROUND TRUTH

Physical measurement

Defect dimensions measured in millimeters against the ground truth established by physical inspection of each sample.

ACCEPTANCE

4 thresholds

Accuracy ≥ 90% · MDR ≤ 10% · FDR ≤ 10% · Per-image processing within real-time conveyor budget.

Statistical Treatment
Metrics we used

All ten formulas that drive the system.

The quantitative evaluation is supported by the confusion matrix and validation loss analysis. Seven derived performance metrics characterize the model and system; three more define the thermal anomaly rule, the soft-voting ensemble, and the composite severity score that produces every verdict.

EQ 1

Precision

Fraction of predicted No-Good classifications that were truly defective. Minimizes unnecessary rejection of conforming sheets.

EQ 2

Recall

Fraction of actual defective sheets correctly detected. Critical for safety — missed defects propagate downstream.

EQ 3

F1 Score

Harmonic mean of precision and recall. Particularly informative under class imbalance (8 Good vs. 42 No-Good).

EQ 4

Accuracy

Overall classification correctness. Primary acceptance criterion — minimum 90% threshold.

EQ 5

mAP@0.5 / mAP@0.5:0.95

Mean Average Precision at IoU 0.50 (lenient) and averaged across 0.50–0.95 (strict). Primary YOLO benchmark metric.

EQ 6

False Detection Rate

Proportion of Good sheets incorrectly flagged as defective. Drives unnecessary rejection cost. Target ≤ 10%.

EQ 7

Missed Detection Rate

Proportion of defective sheets the system missed. Safety-critical — these defects reach downstream. Target ≤ 10%.

EQ 8

Thermal Anomaly Rule

Three conditions hold simultaneously across 3 consecutive frames. Two filters reject sensor noise; persistence rejects transients.

EQ 9

Soft-Voting Ensemble

Weighted average of per-modality confidence. Weights calibrated to per-sensor accuracy (90% webcam / 80% thermal). τ = 0.50 decision threshold.

EQ 10

Composite Severity

0–100 score blending visual detections, thermal anomaly, hotspots, and fusion overlaps. Each component capped so no single sensor dominates.

CHAPTER 03 03

Results

How we solved each objective — with the evidence on the test set.

Dataset Detection Web application Prototype Compliance
Objectives · Refresher
Before we walk through the results

A quick reminder of what we set out to do.

Each result that follows answers one of these five objectives directly — with the specific numbers, artifacts, and evidence that solve it.

Dataset
Acquire and verify a dataset for model training against industry standards.
Detection
Detect scratches and dents through deep learning + suitable camera placement.
Web application
Develop a web application that can be integrated with manufacturing systems.
Prototype
Build a working prototype demonstrating end-to-end defect detection.
Compliance
Evaluate against ISO 9001 / 10012 with accuracy meeting the industry threshold.
Result · Dataset
Objective: dataset acquisition + verification

Dataset validated to 0.904 mAP.

0mAP@0.5

Custom-collected dataset plus NEU-DET supplement trained YOLOv11 to 0.904 mAP@0.5 on the held-out test set — clearing the 90% literature threshold for prototype-level surface defect detection.

0
Annotated images
0
Bounding boxes
350/100/50
Train / Val / Test
0.889
Mean F1 (YOLOv11)
Result · Detection
Objective: detection + camera placement

YOLOv11 selected. Both classes above 0.87 F1.

Three YOLO variants trained under identical conditions. YOLOv11 won on per-class F1 and validation loss. Convergent mounting geometry adopted for both sensors at comparable working distances.

0
F1 — Dent class
0
F1 — Scratch class
0
Inference time per image
0
Val loss at epoch 100
v8

YOLOv8

mAP 0.791 · F1 0.776

v11

YOLOv11 — SELECTED

mAP 0.904 · F1 0.889

26

YOLO26

mAP 0.847 · F1 0.835

Result · Web application
Objective: web app integrable with manufacturing workflow

Steel IRIS deployed at kirabase.net.

Three-tier architecture: Vite/React frontend on Vercel, Supabase PostgreSQL backend with row-level security, Python hardware acquisition layer streaming from the Mini-PC over HTTPS.

0
Functional modules
0
RBAC tiers
3-tier
Stack architecture
RLS
DB-layer access control
MODULES

Live inspection · Manual review · Analytics · Kanban · Steel sheets · Scan logs · Inspections · Equipment · Auth

ROLES

Operator → Inspector → Quality Manager → Admin → Super Admin

Result · Prototype
Objective: end-to-end prototype demonstration

106 ms per cycle. Real-time conveyor inspection.

0end-to-end

Average pipeline latency per inspection cycle — ~9.4 fps continuous operation, well within the 200 ms per-sheet budget. Validated at 12 cm/s conveyor speed under controlled LED illumination.

~9.4 fps
Continuous throughput
12 cm/s
Validated conveyor speed
₱30,555
Total prototype budget
2 sensors
Webcam + MLX90640
Result · Compliance
Objective: ISO 9001 / 10012 + minimum accuracy threshold

96% accuracy. 8 of 12 ISO clauses fully compliant.

0overall

Overall classification accuracy exceeds the 90% threshold by 6 percentage points. Missed-detection rate is well below the 10% safety-critical ceiling. False-detection rate marginally exceeds the 10% target — traced to one surface-oxidation case with a defined mitigation path.

0
MDR (target ≤ 10%)
0
FDR (target ≤ 10%)
8 / 3 / 1
Compliant / Partial / Not
10012 §7.3
Only non-compliant clause
CHAPTER 04 04

Conclusions

How we met each of the five objectives — with the specific evidence.

Dataset (MET) Detection (MET) Web application (MET) Prototype (MET) ISO compliance (PARTIALLY MET) Honest limitations
4.1 · Objective 1
A custom-collected, augmented dataset that meets and exceeds the 90% literature threshold for prototype-level surface defect detection.
How we met it
500 images
Custom-collected on the prototype rig, supplemented with a remapped NEU-DET subset.
682 boxes
Bounding-box annotations across two defect classes (Dent, Scratch) via Roboflow.
350/100/50
Train / Validation / Test partition 425 NG · 75 Good
mAP 0.904
Industry-standard verification — exceeds the 90% literature threshold cited for surface defect detection.
Evidence · §4.1, Tables 8 / 9 / 10
4.2 · Objective 2
YOLOv11 selected after benchmarking three variants. Convergent mounting geometry adopted for consistent capture across both sensors.
How we met it
3 variants
YOLOv8 / YOLOv11 / YOLO26 trained under identical conditions for a fair comparison.
F1 0.877/0.902
YOLOv11 per-class scores for Dent / Scratch — best of the three. val loss 0.754 at epoch 100 · no train/val divergence
9.3 ms
Inference time per image — compatible with real-time conveyor-line inspection.
Convergent mount
Both sensors aimed at the inspection zone at comparable working distances — simplifies ensemble calibration.
Evidence · §4.2, Tables 10 / 11, Figure 4
4.3 · Objective 3
Steel IRIS is live at kirabase.net — full quality-management substrate with evidence trail, role-based access, and approval workflow.
How we met it
3-tier stack
Vite/React on Vercel · Supabase PostgreSQL · Python hardware layer over HTTPS.
9 modules
Auth · Live inspection · Manual review · Analytics · Kanban · Steel sheets · Scan logs · Inspections · Equipment.
5 RBAC roles
Operator → Inspector → Quality Manager → Admin → Super Admin Enforced at the DB row level via Supabase RLS — can't be bypassed client-side.
Kanban approval
Quality Manager sign-off required before any sheet advances — supports ISO 9001 §8.6 release control.
Evidence · §4.5, Figures 9–20, Table 16
4.4 · Objective 4
A continuous-inspection rig integrating both sensors, YOLOv11, thermal evaluator, soft voting, and the web upload — running within real-time budget.
How we met it
106 ms
End-to-end pipeline latency per inspection cycle (~9.4 fps) — well within the 200 ms per-sheet budget.
Mini-PC + ESP32-S3
Separates compute (YOLO + fusion) from sensor acquisition (thermal serial) so neither blocks the other.
₱30,555
Total prototype budget — consumer-grade components, fully reproducible. Thermal cam · webcam · ESP32 · Mini-PC · diffused light · steel samples
Conveyor stop
Physical button + Live Inspection dashboard control — mitigates speed-induced accuracy degradation.
Evidence · §4.4, Figures 4–7, Tables 3–6
4.5 · Objective 5
~
96% accuracy clears the threshold by 6 points. 8 of 12 ISO clauses fully compliant. The remaining gaps are defined engineering tasks, not architectural failures.
Compliance review · click any clause
9001 · 5.3
Compliant
Roles & authorities
9001 · 7.1.5
Partial
Measuring resources
9001 · 7.5
Compliant
Records control
9001 · 8.5.2
Compliant
Traceability
9001 · 8.6
Compliant
Release control
9001 · 8.7
Compliant
Nonconforming control
9001 · 9.1.3
Compliant
Data analysis
9001 · 10.2
Partial
Corrective action
10012 · 6.3
Compliant
Material resources
10012 · 7.1
Partial
Metrological confirmation
10012 · 7.2
Compliant
Measurement process
10012 · 7.3
Not compliant
Measurement uncertainty
Tap any clause above
See how Steel IRIS implements (or doesn't yet implement) that requirement.
Evidence · §4.6, Tables 17 / 18
4.6 · Limits & next steps
Where the boundaries are

Honest limits. Defined fixes.

Every claim in this thesis is bounded by the conditions we actually tested. Every gap has an explicit next step.

L1

Test size n=50

Only 8 Good samples — a single FP moves FDR by 12.5 pts. Fix: expand to balanced n ≥ 200.

L2

Lab vs. factory

Trained at 12 cm/s, stable LED. Real factories have variable lighting, vibration, dust. Fix: site-specific re-calibration before deployment.

L3

Corner detection

Webcam underperforms at sheet corners due to barrel distortion and uneven edge illumination. Fix: lens correction + directional side lighting.

L4

Class coverage

Only two defect classes trained. Pitting, scale, laminar cracks excluded. Fix: dataset expansion + retraining.

L5

Measurement uncertainty

No formal R&R budget. Fix: per-modality repeatability/reproducibility study to close ISO 10012 §7.3.

L6

FDR 12.5% > 10%

One borderline surface-oxidation case (LIVE-003). Fix: fusion veto rule — reject NG when thermal Δ<1°C and YOLO conf<0.55.

CHAPTER 05 05

Demo

It's deployed. You can open it now.

Steel IRIS live at kirabase.net Defense-prep resources
5.1 · Live demo
Steel IRIS · production

It's real. It's deployed.

React frontend on Vercel, Supabase PostgreSQL backend, Python hardware acquisition layer streaming from the Mini-PC. The same nine modules you saw earlier — running live.

Defense panel

Thank you. Questions?

We've prepared backing materials in case the panel wants to dig deeper.

Eugene Gabriel Narvadez
System architect
Jazther Hebrado
Hardware integration
Ana Marie Ramos
Dataset & ML
Lynzy Anne Revilla
Compliance & QA