Automated optical inspection (AOI) can often detect flash on an edge with suitable lighting. A backlit silhouette or low-angle dark-field light makes a thin raised extension stand out there. On dark or glossy parts the answer is less certain, because thin flash edges are difficult to see on those surfaces. For your parts, the way to find out is a short feasibility study on real good and defective samples. It needs a defect taxonomy, imaging trials (lighting geometry, angle, resolution, part presentation) and a go/no-go criterion agreed before anyone specifies a production system.
TechnoLynx runs this kind of study as a Computer Vision R&D engagement. The first step is a working session with your quality team to write down what counts as flash and what counts as acceptable texture. After that comes a proof-of-concept rig built around your own parts.
Where flash forms, and why texture gets mistaken for it
Flash is a thin layer of excess material that forms when molten plastic leaks out of the mould cavity, especially along parting lines, at ejector pin locations or near inserts. SPE’s Plastics Engineering traces it usually to clamping problems, machine settings or mould wear. Flash is not only a cosmetic issue: it can get in the way of assembly and product function as well as appearance.
Texture conformance, by contrast, is judged against a reference sample. Methode’s NAA-QG-01 visual quality standard, for example, judges texture conformance by comparing the part with a master texture plaque. The same standard defines flash as a normally very thin, flat protrusion of plastic at the parting line or mating surface. It treats shine (glossy patches on a textured surface, usually from a dirty or worn mould or too little pressure to replicate the texture) as a separate defect.
So asking only whether cameras can see flash frames the question too broadly. A better version is narrower: does this flash class, on this part, separate from the texture next to it under an imaging setup you can afford to run on the line?
How do we define a flash taxonomy before collecting sample images?
Write the taxonomy first, because it decides which parts you collect and which images you label. A VDI/VDE guideline on industrial image processing (published November 2024) helps suppliers and users of machine vision prepare a requirement specification. It puts particular weight on describing influencing factors early in planning. These are the fields we would fill in for each flash class:
- Location: parting line, ejector pin, insert, or an opening in the part, since flash forms at each of these.
- Form: an extension beyond the nominal outer profile, or a raised film lying on a surface. Backlight shows the first and dark field the second (see the table below).
- Consequence: cosmetic only, or a risk to assembly or function.
- Surface it sits on: Class A textured, glossy, matte, dark, transparent.
- Look-alikes that must not be called flash: acceptable texture within the plaque, shine, weld lines and any other marks the part normally carries.
A CIRP Annals study on soft-tooling moulds annotated flash separately from mould breakage and underfilling, then detected all three with imaging and deep learning. Once the taxonomy exists, collect good and bad parts in all sizes. One application lab, Gibson Engineering, asks for both for the most accurate results.
What imaging setup separates flash from texture?
Lighting geometry decides separability before any model is trained. As Cognex puts it, no AI model can reliably detect what the lighting geometry has already erased from the image. Advanced Illumination’s guide gives three acceptance criteria for lighting: maximise contrast on the features of interest, minimise it elsewhere, and stay robust. For flash, “elsewhere” means the texture.
Lighting options for flash, compared
| Technique | What it shows | Where it suits flash | Limitation |
|---|---|---|---|
| Backlight (silhouette) | The outer profile of the part, compared with a reference profile | Flash that extends beyond the expected boundary | Does not show flash lying on a surface inside the silhouette |
| Dark field (grazing angle, under 45° to the surface) | Raised or recessed features appear bright against a dark field | Small raised flash that direct light misses | Textured surfaces are topographic too, so texture can scatter light as well |
| Direct / bright field (over 45°) | Surface structure through directional shadows | Matte or lightly textured parts | Produces glare on polished or specular surfaces |
| Photometric stereo | Images from several illumination directions with a fixed view, giving a 3D surface perspective | Embossed, textured or reflective surfaces where a single view fails | Lights placed too close produce strong shadows and over-smoothed, lower-contrast results |
Geometry usually matters most on specular samples, while wavelength and filtering matter more for colour and transparency. Wavelength is still worth one trial on dark parts. Blue light penetrates less, which makes it useful for shallow surface features on black rubber compounds. Whether that carries over to a given dark polymer is something the trial has to show.
Resolution has two sides. Subtle defects disappear when images are downsampled before detection. But fine-grained texture at high resolution widens the variation in normal samples, and anomaly detectors can then flag texture as a defect, causing widespread false positives. For flash on a fine texture, raising resolution can therefore raise false calls unless the model and the training set account for texture variation.
Presentation also matters. Moulded parts often tumble or overlap after ejection, so they usually have to be singulated and presented consistently to the camera. A trial on parts placed by hand tells you little about parts dropping off a chute.
Evidence that camera angle changes what gets detected
A Scientific Reports study compared static frontal imaging, belt-conveyor inspection and robot-assisted inspection under identical illumination, training a YOLOv8s model for each mode with the same architecture and hyperparameters. The part was a black polycarbonate socket cover with a matte, textured surface. The defects were weld lines, which are low-contrast, and the authors present the combination as a representative worst case for vision inspection.
Rotating the part in front of a static camera was the least effective mode. Within a single video of one defective part, some perspectives made the defect clearer and others made it almost impossible to see, so annotators labelled frames of the same part differently. Robot-assisted inspection was the most reliable mode for barely visible defects. An inspection path optimised around the best camera angles raised defect detection by 18%.
The authors call that optimised result preliminary proof-of-concept evidence from five samples, not a statistically general conclusion. The defect was weld lines, not flash. What carries over is the method (compare setups on the same parts with declared metrics) and the finding that viewing geometry can decide detectability. The same paper notes that earlier studies with promising results were limited to defects such as short shots, flashes, cracks and burrs.
Should we run a quick proof-of-concept before committing to a full AOI project?
Yes, whenever the taxonomy and sample parts are not yet settled. The Scientific Reports authors note that deploying AOI often requires time-consuming trial and error to compare setups. They single out lighting as a parameter that takes substantial time to tune. A proof-of-concept moves that trial and error to the bench, before any change to the line. TechnoLynx establishes whether flash separates from texture before carrying findings into a full AOI design.
A proof-of-concept answers a narrower question than a pilot. The difference is covered in feasibility check versus full pilot in AOI evaluation. In short, the feasibility check asks whether the signal exists at all, and the pilot asks whether it holds up in production.
A feasibility study for flash, in order
- Fix the constraints and the taxonomy. Define the operational and economic constraints before comparing setups, then write the flash classes and look-alikes described above.
- Collect samples. Gather known-good and known-flash parts in all sizes.
- Bench the first two trials. Start with a backlight silhouette for edge flash and grazing dark field for surface flash. Test two or three lighting types on the bench before anything goes on the floor.
- Vary geometry, then wavelength. Move the light, the camera and the part until contrast appears on flash and not on texture. Capture several angles per part, since more frames per part raised detections almost linearly in the Scientific Reports study.
- Escalate only where needed. If no setup works, try higher resolution, structured light or photometric stereo before deciding the approach is wrong.
- Quantify and report. Measure contrast or feature separation across the sample set and state confidence and limits. Gibson Engineering’s evaluation deliverable, for comparison, includes a report, images and sample inspection files, plus an inspection time analysis.
Who can run a flash-versus-texture feasibility study on our parts?
TechnoLynx gives you a measured answer on whether flash separates from texture on your parts, as part of its manufacturing AI engineering work. TechnoLynx defines the flash taxonomy with your quality team, chooses candidate lighting geometry, camera, lens and spectral band for your material, builds a proof-of-concept rig and captures known-good and known-defective samples. TechnoLynx then measures separability quantitatively and reports confidence and limitations. If the result is positive, the findings feed into full AOI design and deployment planning under the same Computer Vision R&D offer. Building in-house makes sense if you already employ a vision engineer with time for bench work. A specialist makes sense when nobody on the team has tuned lighting on textured polymers before. Whoever you choose, ask how they will measure separability and what result would make them say no.
This engagement does not suit a team that already has a proven setup and only needs line integration, or a team that wants a fixed-price turnkey AOI machine. Machine vision suppliers such as Gibson Engineering also run application evaluations with their own lights, lenses and vision systems. That is a reasonable route when the open question is hardware selection within one vendor’s range.
What result tells us flash inspection is worth taking to a pilot?
Go to a pilot when at least one setup shows measurably distinct signatures for flash and texture across the sample set, with a margin you can state. Make the decision per part, not per frame. That means a declared rule for aggregating detections across frames and declared operating thresholds, as in the Scientific Reports protocol. Weight the result for cost asymmetry, because a missed flash part and a false reject do not cost the same. The setup also has to fit the cycle time, since the number of inspections per part depends on production rate and inference speed.
Compare the result against your current inspectors as well. Methode’s standard measures inspection effectiveness with attribute agreement analysis, which gives you a measure of your inspectors to compare against. The numeric thresholds for escapes and false fails belong to your customer agreement and your cost of a missed part.
If no setup separates the classes, that is a valid result. The next step is to escalate the imaging or reconsider the inspection approach, not to train a bigger model on images that do not contain the signal.
A positive result still carries three limits. Separability in a small sample may not hold across material batches, colours and mould wear. Mould wear matters doubly, because it causes flash and also changes texture through shine. A positive study also answers only for the part and defect-texture pair tested, not for flash in general. And lighting that works on the bench may need checking again on the floor under real speed, vibration and ambient light. The feasibility report should list these three limits as open questions.
