CLOSE
aigocy@gmail.com

CUP

TWITTER (X) DRIBBBLE LINKEDIN

Configuration

COLORS
CUSTOM CURSOR
Skip to main content

espialtech

Menu
Start a Project

CastSight

Home
/
Blog
espial
July 27, 2026
No Comments
CastSight
X-ray Inspection

Real-time detection of internal flaws in aluminium castings on the production line,
using deep learning trained on defects that never physically existed.

Project Details

Porosity, shrinkage and inclusions inside an aluminium casting are invisible from the outside and expensive downstream — a flawed part machined, assembled and shipped costs orders of magnitude more than one caught at the source.

CastSight inspects castings on-line with an industrial X-ray tube (160kV-225kV) and a flat-panel digital detector, positioned by a motorised rotary stage that captures the part from multiple angles for volumetric coverage. Conveyor-integrated proximity sensors and PLC triggers fire the capture automatically. Detection runs on an NVIDIA Jetson AGX Orin at 18 FPS, in line with production speed, and drives an automated reject mechanism with full batch traceability and defect heatmaps for the quality team.

DELIVERABLES
X-ray vision pipeline Synthetic data generation Edge inference
INDUSTRY
Automotive Manufacturing

Project Research

The hard problem was training data. Real internal defects are rare by definition — that is the point of the quality programme — and annotating X-ray imagery requires a specialist.

So we generated the defects instead: a 3D ellipsoidal simulation model that synthesises realistic porosity and inclusion geometry and maps it onto clean castings with physically plausible attenuation. Thousands of synthetic defects, zero manual annotation. We then ran a multi-model evaluation — YOLOv3, YOLOv5, RetinaNet and EfficientDet — against a held-out set of real flawed parts to confirm the synthetic-to-real transfer held up.

Project Results

CastSight achieves a 98% flaw detection rate at 18 FPS, fast enough to inspect every part rather than a sample. Manual annotation was eliminated entirely — thousands of synthetic defects were generated instead — and the line reject mechanism now operates automatically with batch-level traceability, so a quality escape can be traced back to the shift and the mould that produced it.

SiteGuard
Safety Analytics
View Project

SiteGuard
Safety Analytics

DESCRIPTION
On-premise video analytics for PPE compliance, intrusion detection and zone safety — 20 FPS tracking, running as a single executable with no cloud, no servers, no internet.
DELIVERABLES
Open vision models, On-premise deployment, Automated compliance reporting
INDUSTRY
Industrial Safety & Compliance