Artificial intelligence (AI) is reshaping machine vision at speed. Deep learning tools that once required specialist expertise are now accessible to vision engineers through mature software frameworks and pre-trained models. The results are genuinely impressive: systems that detect subtle defect classes, adapt to product variability, and improve over time as more data becomes available.
But, as AI has become more prominent in machine vision, a worrying assumption has taken hold. Increasingly, AI is treated as a magic layer, a sufficiently powerful algorithmic stage capable of correcting whatever the imaging system fails to capture cleanly. End users request AI-based inspection without specifying image quality requirements. System integrators face pressure to deploy AI solutions on timescales that leave little room for illumination optimisation. The implicit belief is that the algorithm will manage.
"We see this regularly," says Alexandre Cottereau, Head of Product & Marketing at Effilux. "There is a growing expectation that AI will solve imaging problems that should have been addressed in hardware. By the time the limitation becomes apparent, the project is already well advanced and the cost of going back to optimise the illumination is significant."
It will not simply manage. And understanding why begins with a simple principle: the best AI model in the world cannot recover information that was never captured by the camera.
What the algorithm actually receives
A machine vision AI does not perceive the physical world. It processes pixel data, a numerical representation of the light that reached the sensor during image acquisition. Everything the model knows about the inspection target is encoded in that data. Features that are not represented in the image cannot be detected, classified or segmented, regardless of how sophisticated the network architecture is.
Consider a common industrial challenge: detecting micro-scratches on polished, cylindrical metal components. Under standard diffuse white light, specular reflections wash out the metal surface, rendering the scratch virtually invisible to the sensor. If an engineering team attempts to train a deep learning model on these images, the network will struggle, generating high false-negative rates regardless of how many thousands of sample images are fed into the training pipeline.
Place that same component under darkfield, low-angle directional illumination, and the flat surface reflects light away from the lens while the scratch scatters light directly into the camera. The defect instantly appears as a high-contrast white line on a dark background. The AI model can now identify the defect instantly, with almost 100% confidence and minimal training data.
"AI starts with photons," Cottereau states. "Everything the algorithm knows about the part in front of the camera comes from the light that reached the sensor. If that information is missing or degraded, the model cannot reconstruct it. That is simply not what neural networks do."
The real cost of poor image quality
In practice, the belief that AI can compensate for weak imaging tends to manifest as a sequence of escalating interventions. Initial model performance is disappointing. The response is to collect more data, extend training, and increase network complexity. Each step may yield marginal improvement while the root cause, inadequate image quality, remains unaddressed.
The downstream costs accumulate. Models trained on poor-quality images require larger datasets, because the network must learn to extract signals from data where the signal-to-noise ratio is inherently poor. Training cycles lengthen. Models become brittle, sensitive to changes in imaging conditions that a well-specified lighting system would render inconsequential. False positive and false reject rates remain stubbornly high.
"Sometimes a better light is cheaper than a better AI model," Cottereau observes. "We have seen projects where months of additional development could have been avoided by addressing the illumination at the start. The investment in a well-designed lighting system is almost always modest relative to the cost of what follows when image quality is poor."
Where illumination sits in the imaging chain
The imaging chain in a machine vision system runs from object through illumination, optics and sensor to algorithm and decision. Information can be lost or degraded at each stage. It cannot be recovered once lost.
Illumination occupies a uniquely influential position in this sequence. The choice of lighting geometry, wavelength and intensity distribution determines which features of the inspection target generate contrast and which do not. A defect invisible under diffuse illumination may be immediately apparent under directional or darkfield lighting.
A surface contamination indistinguishable under white light may be clearly visible under UV fluorescent excitation or specific short-wave infrared (SWIR) bands. These are not marginal differences. They are the difference between a detectable and an undetectable defect, independent of whatever algorithm is applied downstream.
"Illumination determines what information enters the system," Cottereau explains. "Contrast, signal-to-noise ratio, feature visibility, repeatability across shifts, all of these are established at the lighting stage. AI performance starts long before data reaches the neural network. The model is trained on the images the system produces, and if those images do not reliably encode the features that matter, there is nothing downstream that will fix that."
Designing for AI from the start
The practical implication is straightforward: AI-ready machine vision systems must treat image quality as a primary design objective, not a secondary consideration.
This means defining image quality requirements before specifying hardware, and validating illumination performance before collecting training data. A dataset collected from a poorly configured imaging system encodes the artefacts of that system alongside the characteristics of the inspection target.
Models trained on such data become dependent on those artefacts, making them fragile to any subsequent change in the imaging setup.
"Improving image quality is often the fastest way to improve AI performance," says Cottereau. "We encourage engineers to think of illumination as a data-quality tool rather than simply a source of brightness. The goal is to maximise the contrast of the features the algorithm needs to detect, and to eliminate variability that is irrelevant to the inspection decision. If you do that well before you start collecting training data, the model development becomes significantly more straightforward."
Hardware and software should be developed iteratively, with imaging system performance and model performance assessed together. Where illumination changes improve image quality, the effect on model performance should be measured. Where model performance falls short, the imaging system should be the first element reviewed.
AI and lighting as partners
None of this is an argument against AI in machine vision. The capabilities that deep learning brings to industrial inspection are real and still expanding. The argument is that those capabilities will only be fully realised when they are combined with imaging systems designed to give them the information they need.
"The future of machine vision will not be driven by AI alone," Cottereau concludes. "It will be driven by the combination of better illumination, better optics, better sensors and smarter algorithms, each developed with awareness of the others. The companies that understand this will build systems that consistently outperform those that treat AI as a substitute for good imaging."
Machine vision performance is ultimately determined by the quality of information entering the system. As AI models become more capable, the value of high-quality image data increases rather than decreases.
AI starts with photons. The systems built on that understanding will outperform those that are not.
