AI-assisted rail-surface inspection · research-to-pilot technology

InfraVision RailSenseFrom rail imagery to measurable evidence.

A quantitative evidence layer for engineering review: localized rail-surface findings, mask-based geometry, rail-head position, review prioritisation, section-level condition evidence and archive-ready inspection records.

Discuss a validation pilot →Explore locked-run evidence
RailSense inspection finding showing gauge-corner and head-check evidence
LocalizedPIXEL-LEVEL EVIDENCE
MeasuredMM · AREA · POSITION
TraceableCASE RECORD · REVIEW PATH
86images processed
LOCKED V9 EVIDENCE SET
214candidate detections
BEFORE INTERNAL CHECKS
198reported findings
AFTER QUALITY CHECKS
194review-required candidates
CONSERVATIVE TRIAGE GATE
Allcandidate cases retain traceability artifacts
MASK / CHECKLIST EVIDENCE IN THIS RUN
What the layer adds

Beyond detection: a quantitative evidence chain.

The differentiating layer begins after a visual anomaly is found: where it lies on the rail head, how its visible extent is quantified, how the evidence is documented, and how cases are prioritised for engineering review.

01 · DETECT

Localized surface findings

Rail-surface candidates are delineated and retained as reviewable visual evidence rather than reduced to a class label alone.

02 · MEASURE

Physical geometry

Length, width and area are estimated in engineering units when image calibration is available, enabling comparable defect-level records.

03 · CONTEXT

Rail-head zoning

Running band, gauge shoulder and gauge-corner context are carried into the record to distinguish defects that look similar but occur in different locations.

04 · TRIAGE

Review prioritisation

Evidence is ranked so inspectors can see which images and instances deserve attention first. It is a review aid—not an autonomous safety decision.

05 · AGGREGATE

Section-level condition evidence

Defect evidence can be aggregated into section-level condition views and standards-aware indices, with operator-specific calibration during a pilot.

06 · AUDIT

Traceable case records

Mask source, image, geometry, zone, evidence cues and review state are retained so outputs can be audited rather than treated as black-box scores.

Evidence path

One image, three progressively richer views.

The same rail crop is carried from source image to rail-head context and finally to measured findings. Every displayed image below is stored locally inside this website package.

Raw rail-head crop
STEP 01 · SOURCE

Raw rail image

The starting inspection crop is preserved as source evidence.

Rail-head zone overlay
STEP 02 · CONTEXT

Rail-head zone

Running-band and gauge-side context are added before the finding is interpreted.

Measured RailSense finding
STEP 03 · EVIDENCE

Measured findings

Finding class, geometry, priority and review information become an auditable case record.

Locked V9 evidence gallery

Different defect patterns, one consistent evidence format.

These cases were selected to show distinct operational situations—not only the most visually dramatic defects. Click any card to inspect the full-resolution image.

Gauge-corner Head Check case
GAUGE CORNER · MIXED FINDINGS

Head Check / gauge-corner review

Three localized findings, two requiring review; illustrates gauge-side engineering context.

Critical triagemax DSI 0.896
Head Check and Spalling case
GAUGE CORNER · TWO DEFECT TYPES

Head Check + Spalling

Two finding types on the same image, demonstrating multi-defect evidence rather than one-label classification.

Critical triage2 findings
Corrugation and Squat findings
RUNNING BAND · MULTI-DEFECT

Corrugation + Squat

Five localized findings in the running band; useful for demonstrating defect burden and review prioritisation.

Critical triage5 findings
Multiple Wheelburn findings
RUNNING BAND · HIGH COUNT

Many Wheelburn findings

Fifteen small findings demonstrate why count alone should not be confused with severity or maintenance priority.

Low individual priority15 findings
Shelling and Wheelburn findings
MULTI-ZONE · MIXED DEFECTS

Shelling + Wheelburn

Shows how class and rail-head location can coexist in the same review record.

4 findingszone-aware
Single squat finding card
INSPECTION CARD · SINGLE FINDING

Compact squat review card

A simple inspector-facing card view with the localized defect, image location and review suggestion.

Low priority exampleArea 326 mm²
Multiple finding inspection card
INSPECTION CARD · MULTI-FINDING

Mixed wheelburn and squat card

A denser example showing multiple findings on one image, useful for triage workload and review-board design.

4 findingsMixed priorities
Manager view

Move from isolated detections to a reviewable network story.

The locked run contains both instance-level evidence and aggregated triage views. The following summary uses only this single run so numbers are not mixed across experiments.

Priority distribution

198 post-check findings. The labels below represent prototype triage priority, not certified safety risk.

Low
114
Medium
23
High
8
Critical
53

Two RSQI population views

The locked report intentionally carries both numbers and warns against quoting either one without its population definition.

50.0/100 · assessed sections
91.5/100 · full processed set
Interpretation matters: full-set scoring includes processed images with no reported defect as clean. Section-assessed scoring is a different population. Both must be reported together.

Risk / review matrix

Existing engineering risk logic can be used as a review scaffold, but thresholds and consequence classes must be calibrated with the operator.

Rail infrastructure risk assessment matrix

Traceability chain

A structured finding can retain evidence from multiple stages rather than only the final prediction.

RailSense traceability evidence example
  • Source image and localized finding.
  • Rail-head zone / positional context.
  • Mask-based geometry and measured area.
  • Evidence-memory retrieval and manual feature checklist.
  • Review state and proposed maintenance-priority class.
Exploratory scenario views

From current evidence to testable deterioration hypotheses.

These views show how the current defect mix can be translated into traffic- and time-based planning scenarios. They are deliberately labelled as hypotheses for operator validation, not as operational forecasts.

Traffic-scaled Markov envelope showing RU probability

Traffic-scaled Markov envelope

Explores the share of active defects reaching the RU state under slow, base and fast exposure assumptions. The transition model requires repeated inspections of the same assets before it can be treated as a validated forecast.

Visual RSQI proxy deterioration scenario

Visual RSQI scenario

Shows an image-based surface-condition trajectory over time. It is a decision-support research view; local calibration, repeated inspections and operator maintenance records are required for production use.

Important: scenario outputs are not autonomous maintenance recommendations, guaranteed failure times or validated savings estimates.
Trust boundary

Industrial credibility starts with what is not yet proven.

This prototype is designed as an AI-assisted inspection and evidence-triage layer. It is not certified for autonomous maintenance decisions.

What this locked run demonstrates

  • Traceable visual evidence and review prioritisation.
  • Mask-based measurement and rail-head positional context.
  • Section-level surface-condition views and standards-aware mapping.
  • Archive-ready evidence cards for inspector review.
  • 100% traceable output, SAM2-mask traceability and memory retrieval within this run.

What requires operator validation

  • Precision, recall and dangerous false-negative rate on a blind expert-labelled set.
  • Class-confusion performance under the operator's image distribution.
  • Measurement agreement against expert or physical reference measurements.
  • Local mapping from evidence severity to inspection / maintenance action.
  • Repeated same-asset inspections before deterioration forecasting is treated as validated prediction.
One evidence layer, different starting points

Designed to complement the inspection stack you already operate.

The same product narrative works whether an organisation already collects rail-head imagery or needs a small image-acquisition pilot to evaluate the visual layer.

IF YOU ALREADY HAVE IMAGERY

Add the evidence layer.

Run RailSense on existing inspection imagery without replacing geometry, ultrasonic or other condition-monitoring systems.

  • Evaluate on your camera characteristics and operating conditions.
  • Compare the review queue with current inspection workflow.
  • Test whether geometry, zoning and case traceability add useful information.
IF YOU DO NOT HAVE SYSTEMATIC RAIL-HEAD IMAGERY

Start with a controlled visual pilot.

Scope a short acquisition campaign, then evaluate the same quantitative evidence chain on a known section.

  • Define camera setup and rail-head calibration.
  • Capture a limited, traceable section.
  • Compare visual findings with existing inspection records and expert assessment.
Industrial validation pilot

Validate the evidence on your own inspection context.

The next step is a bounded technical validation exercise with shared evidence, expert reference labels where available, and pre-agreed evaluation criteria.

Discuss a pilot → info@railvisionai.com
AI-assisted inspection triage · Not certified for autonomous maintenance decisions
  1. Scope: agree a representative image set or short controlled section.
  2. Blind evidence: obtain expert labels and, where possible, reference measurements.
  3. Run: process the data without tuning to the evaluation labels.
  4. Compare: precision / recall, false negatives, class confusion, measurement agreement and review workload.
  5. Decide: determine whether the evidence layer adds enough operational value to scale.