Skip to main content

The Biggest Enemy of AI in Machine Vision: Physics

How object geometry, material properties and lighting influence automated inspection

Today’s article focuses on the question of how strongly the geometric and optical properties of the inspected object itself determine the design of a machine vision system – in other words, the substrate on which the features appear.

In practice, this is often where it is decided whether an inspection will work robustly or fail due to basic physics.

How object geometry affects machine vision performance

Substrate geometry: flat, structured or curved

The geometry of the object plays a central role:

  • Flat, even surfaces can usually be captured relatively easily using conventional lighting configurations such as coaxial, dome or area illumination.
  • Ribbed, structured or machined surfaces can produce strong intensity variations depending on the direction of the light. These variations may obscure actual defect contrasts or even imitate defects.
  • Curved or strongly three-dimensional components create continuously changing angles of incidence and reflection, resulting in local overexposure, hotspots and shadows.

All of these factors directly influence the positioning of both the camera and the illumination.

    The best illumination strategies for different surface geometries

    • For flat surfaces, the optical axis can often be positioned almost perpendicular to the object. The illumination should then be selected according to whether reflective or diffuse components need to be emphasised, for example using coaxial or dome illumination.
    • For ribbed or structured surfaces, combining several lighting directions is often useful to reduce directional effects. Alternatively, deliberately angled illumination can highlight certain edges while suppressing others.
    • For curved components, a small point-like light source is almost always problematic. Large light sources that surround as much of the object as possible — such as dome lights, large area lights or large-diameter ring lights — are generally more robust. Multiple or moving illumination positions may also be required.

    For AI-based applications in particular, it is essential to prevent purely geometry-related artefacts, such as hotspots on curved surfaces, from being learned as defects. A lighting concept that takes the object geometry into account is therefore not optional. It is fundamental to reliable inspection.

    Optical properties: Matte, reflective or transparent

    In addition to geometry, the optical properties of the substrate determine how light interacts with the object:

    • Matte, diffusely reflecting surfaces scatter light relatively independently of the angle of incidence. Contrast is mainly created by differences in reflectivity.
    • Reflective or glossy surfaces follow the law of reflection much more closely. Even small changes in angle can cause major changes in image intensity.
    • Transparent or translucent materials add transmission, refraction and volumetric scattering to the equation.

    These properties have a direct impact on the choice of illumination.

    Glossy surface

    Brushed surface

    Matte surface

    Selecting the right illumination for different materials

    • Matte surfaces: The goal is homogeneous and stable illumination with a high signal-to-noise ratio. Area lights, diffusers and, where necessary, polarisation filters on both the light source and the lens can help reduce unwanted residual reflections.
    • Glossy surfaces: A conscious decision must be made as to whether the specular reflection should be used or suppressed. Coaxial illumination is suitable for presenting flat, mirror-like surfaces evenly and brightly, while defects such as scratches or residues create contrast. Flat-panel or dark-field illumination, on the other hand, can emphasise edges, scratches and topographical defects.
    • Transparent or translucent surfaces: Backlighting is often useful for making edges and internal irregularities visible. For surface defects, angled or coaxial illumination may be more suitable, possibly combined with polarisation to control disruptive reflections from the surroundings.

    The camera position is closely linked to the material properties. When specular reflection is deliberately used, the camera should be placed in the corresponding reflection direction. If reflections are unwanted, the optical axis should be positioned outside these angles.

    Homogeneity vs. variability: Corrosion, ageing and environmental influences

    Applications become particularly challenging when the substrate itself varies significantly — either locally within one object or globally between different objects.

    Typical examples include:

    • Corrosion on metal components stored outdoors
    • Weathered railway sleepers
    • Surfaces affected by dirt, moisture or biological influences

    Several effects overlap in these situations:

    • Geometric changes such as chipping, increased roughness or wear
    • Optical changes such as colour shifts, altered surface roughness and changing reflection properties
    • Significant inhomogeneity within an individual object as well as between different samples

      Why consistent illumination is essential for AI models

      • Perfectly homogeneous reference images are practically impossible under these conditions.

        The system must therefore be designed to handle normal variations as robustly as possible.

        • Large, diffuse light sources such as dome or area lights help prevent local extreme reflections and average out microscopic surface irregularities.
        • A carefully selected angle of incidence can ensure that rough or corroded structures remain visible without allowing every patch of corrosion to dominate the image when the actual inspection is focused on different defects, such as cracks or structural damage.

        Long-term stability is another important factor:

        • Outdoor systems are exposed to changing ambient light, dirt and moisture.
        • Shielding against ambient light, robust enclosed illumination units and regular recalibration — potentially using reference targets — are essential.

      AI models benefit significantly when illumination represents the normal variability of the substrate consistently. This allows genuine anomalies to be detected reliably, even when the surface conditions vary considerably.

      Camera and illumination arrangement: The most important system design decision

      All of the aspects described above ultimately come together in one central decision: the position of the camera and the illumination within three-dimensional space.

      Different configurations are required depending on the surface and the type of defect:

      • On-axis vs. off-axis: Is the light introduced along the optical axis using coaxial illumination, or are angled light sources used?
      • Ring, line, area or dome illumination: Should the light be highly directional or distributed across a wider area?
      • Static vs. sequential configurations: Is one lighting configuration sufficient for each image, or are several lighting conditions required in sequence to reveal different defects and surface effects?

      Multiple illumination setups for robust AI inspection — better data for better decisions

      For highly variable or complex substrates, it is often useful to:

      • Capture the scene using several light sources positioned at different angles
      • Generate multiple images using different illumination configurations, such as a combination of dark-field, coaxial and dome illumination
      • Combine these images into a derived multi-illumination image, for example by processing them as separate channels or by using maximum, minimum or weighted averaging operations
      • Feed this combined image into the AI model for evaluation

      This makes the inspection less sensitive to local outliers such as hotspots and shadows.
      It also helps both conventional and AI-based methods distinguish relevant intensity patterns from those caused exclusively by material and lighting effects.

        Why understanding physics is essential for AI-based machine vision

        A careful analysis of the geometric and optical properties of the substrate is not a theoretical luxury.
        It is the foundation of every robust machine vision solution.
        Anyone who skips this step and simply selects a camera risks preventing both
        conventional algorithms and AI models from ever reaching their full potential.

        FAQ – Machine vision, AI inspection & illumination

        Why do AI inspection systems fail despite using good models?

        In many cases, the problem is not the model but the image quality.
        When illumination, geometry and material properties are not properly controlled, the system learns unstable or irrelevant features. Even the best AI cannot compensate for poor physical conditions.

        Which type of illumination is best for machine vision?

        There is no universal solution.
        The ideal illumination depends on the geometry and material of the object. Flat surfaces often work well with coaxial or diffuse illumination, while structured or curved components usually require multidirectional or large-area lighting.

        Why are reflective or glossy surfaces so difficult to inspect?

        Small changes in angle can cause major variations in the intensity of reflected light.
        This leads to glare, bright hotspots and unstable contrast, which can confuse both conventional algorithms and AI models.

        Can more training data compensate for poor lighting conditions?

        Only to a limited extent.
        More data may help, but it cannot fully compensate for inconsistent or physically inadequate image acquisition. Improving the illumination setup is often significantly more effective than simply increasing the size of the dataset.

        How can I make my AI inspection system more robust?

        Focus on consistent, physically sound image acquisition.
        Use stable illumination, reduce unwanted reflections and design the system so that it accounts for natural variations in materials and surfaces. In many cases, combining several illumination strategies can significantly improve reliability.

        Your AI Is Only as Good as Your Images

        When an inspection system struggles with false positives, unstable results or poor generalisation, the underlying cause is often the physical setup rather than the algorithm.

        We help companies develop machine vision systems that perform reliably under real-world conditions by aligning illumination, camera positioning and material behaviour from the very beginning.

        Get in touch to discuss your application and turn your inspection into a robust, scalable solution.