Two real industrial systems built around the same engineering principle: reconstruct the measurable state of the physical product, compare it with a calibrated or ideal state, and make the production decision from the deviation.
The projects below solve different problems — gradual PET-preform color drift and surface defects in a moulded plastic product — but both treat computer vision as part of the production process rather than as an isolated recognition demo.
01 · CYCLOPE-COLORReal PET-preform inspection installation with controlled illumination and operator workstation.02 · SURFACE DEFECT INSPECTIONOptical inspection cell for a moulded plastic lampshade.
01Acquire under controlled conditions
02Reconstruct measurable product state
03Compare with calibrated / ideal state
04Signal or production decision
01 · PET PREFORMS
Cyclope-Color
Real-time color inspection and statistical quality control for PET preforms.
Small color changes can accumulate gradually and may be difficult or impossible for a human operator to notice reliably. Cyclope-Color was built to detect those changes continuously during production.
The system was tested across four PET-preform categories: transparent colorless, colored but visually transparent, dark transparent, and colored opaque products.
OBJECT LOCALIZATION→RELATIVE COLOR MEASUREMENT→PROFILE COMPARISON→TOLERANCE VIOLATION→STATISTICS / ALARM
Four-channel controlReal-time inspection uses mean gray plus the L, a and b coordinates of LAB color space. The project materials report discrimination of color differences down to about 3%.
Adaptive profile calibrationA profile-setting module controls illumination and camera exposure, then stores product-specific compensation functions and tolerances for gray, L, a and b.
Production monitoringThe online module performs continuous testing, keeps local history, logs operations and signals unacceptable deviations from the active profile.
Real-time monitoring by four tolerance channels: gray and LAB.Local production statistics: gradual drift remains visible instead of being reduced to a single pass/fail result.
Object localization also creates a path toward surface-defect and geometry analysis.
The system does not merely ask “what color is this object?” It continuously measures whether the production process remains inside the calibrated color state.
02 · PLASTIC SURFACE
Lampshade Defect Detection
Automated optical inspection of the IKEA MELODI plastic lampshade inside the production cycle.
The task had previously been performed manually by a human operator. The system was designed to detect scratches and dents, surface deformations, finish defects, perforations, tears and holes, and dark/light or speck defects.
Here computer vision is not a separate workstation at the end of the process. The inspection result is part of the production logic and is returned directly to the robot handling the product.
MOULDING→ROBOT LOADS PRODUCT→OPTICAL INSPECTION→GOOD / BAD / UNRESOLVED→ROBOT ROUTING
The system is integrated into the production cycle: the robot places the product into the inspection cell and receives the result.
ANALYTICAL CORE
Ideal optical model → deviation → defect.
The core approach constructs an ideal optical model of the current product in real time and calculates deviations from that model. The inspection therefore does not depend only on a fixed catalogue of defect appearances.
The analytical output includes a reconstructed surface model, configurable tolerance analysis, a defect list and explicit localization of detected defects on the surface.
Example output: surface 3D model, tolerance module, defect list and defect marking.
CAMERASBasler area scan
CONTROLModBus
SOFTWARELinux · Python 3
STACKNumPy · SciPy · OpenCV
The common principle is not to recognize an image, but to reconstruct the measurable state of the physical product and make the production decision from that state.
