AI Vision Lighting Best Practices

The right lighting depends on what you need to detect. Diffuse dome lighting works well for surface defects on matte materials. Directional or dark-field lighting emphasizes surface topography, making scratches and raised features stand out. Coaxial lighting is preferred for reflective metals and polished surfaces. LED ring lights are common for general-purpose inspection. Testing multiple configurations on your actual parts is the most reliable way to identify the best setup.

Lighting directly controls contrast between defects and background. Poor or inconsistent lighting forces the AI model to compensate for image quality issues rather than focusing on true defects, which increases false positives and false negatives. Consistent illumination across all inspected parts is more important than absolute brightness level.

Fixed lighting is strongly preferred for production inspection. Variable lighting introduces inconsistency that degrades model performance over time. Once you identify the optimal configuration during setup, lock it in and shield the inspection station from ambient light changes. Adjustable setups are useful during initial configuration and testing, but should be fixed before going into production.

Reflective surfaces require careful lighting geometry. Structured or polarized lighting helps reduce specular highlights that obscure defects. 2.5D lighting techniques used in systems like the OV80i combine multiple light angles to create consistent contrast on metal, glass, and coated surfaces. Telecentric lenses paired with on-axis illumination also reduce the impact of surface reflectivity.

Existing factory lighting is generally not suitable for precision inspection. Overhead fluorescents and general area lighting create shadows, hotspots, and vary significantly with time of day and nearby activity. Dedicated, enclosed inspection stations with controlled illumination produce far more consistent results and reduce the training data needed to achieve target accuracy.