Tire classification with deep learning: capture and validate

Tread classification starts with image acquisition, not a neural network. A model cannot reliably distinguish fine tread features when wheel rotation, reflections or poor viewing geometry erase them. The practical task is to capture repeatable tread crops, establish trustworthy labels and control the consequence of a wrong decision.

This application is for teams designing a controlled vehicle inspection lane. It connects camera, lighting and trigger choices to classifier validation. The output is a supported tire-type classification or a referral for manual inspection—not a tread-depth measurement, traction prediction or roadworthiness approval.

Define the decision before collecting images

Start by defining what the system is allowed to decide. Summer, winter and all-season categories are not necessarily separable from every tread image. Different categories can share visual features, while wear, repairs and contamination change their appearance.

Create a written labeling policy using qualified inspection and any identification information required by the applicable rules. Do not assign ground truth solely from the same tread appearance that the model will evaluate. Where the category cannot be established independently, exclude that sample from definitive class labels or assign an explicitly defined unresolved category.

Keep three outputs separate:

A missing wheel detection is incomplete coverage, not evidence of an acceptable tire. Track each required tire position and distinguish a tire-level result from a vehicle-level result.

For terminology around acquisition, processing and decisions, use the machine vision terminology guide. Establishing these boundaries early prevents a classification demonstration from becoming an unsupported safety claim.

Arrange the view around the tread band

Position the camera to see exposed tread, not mainly the sidewall or wheel rim. An oblique view can reveal tread blocks and narrow grooves, but it also changes apparent feature size across the image. Verify visibility on representative wheel sizes, steering angles and body clearances before selecting a lens.

A useful crop needs enough tread context to distinguish patterns. A close view of one groove may resolve detail while losing the arrangement that identifies the tread family. Conversely, a whole-vehicle image may contain too few pixels across the tire.

Use area scan cameras to capture a defined inspection zone. The camera finder helps narrow interface and shutter choices, but final selection also requires actual image dimensions, trigger behavior and lens compatibility.

Prefer global exposure for moving tread. Compare the global-shutter MV-CU020-90GC with the rolling-shutter LS-CMR-0166 as examples of different acquisition architectures. A rolling shutter requires additional assessment of row timing, geometric distortion and illumination compatibility.

All 2 share: Sensor size 1/1.7". Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Do not assume that a short exposure eliminates rolling-shutter distortion. Different rows can still represent different wheel positions. The moving-part exposure guide explains the distinction between freezing local motion and selecting a shutter architecture.

Plan the station so commissioning and maintenance do not require anyone to approach moving wheels. Traffic management, guarding and any safety-related controls require their own qualified design and validation.

Calculate pixels and motion blur together

Choose a minimum feature size from representative tread samples. Then decide how many pixels are needed across that feature for repeatable recognition. This is a sampling target, not a promise that the classifier will distinguish categories.

For an illustrative design, suppose the tread crop spans 200 mm and the smallest feature of interest is 0.5 mm. Requiring 4 pixels across that feature gives:

Check this requirement against the usable image width after cropping and perspective correction, not just total megapixels. The image downscaling guide explains why resizing later cannot recover detail that acquisition missed.

Now consider motion. Assume a vehicle speed of 18 km/h:

For a wheel rolling without slip, a tread point can reach twice the vehicle speed relative to the ground. Actual blur depends on the projected motion in the camera image. Use that projection for the final design; the maximum provides a conservative starting estimate.

Suppose the selected image crop contains 2000 pixels across 200 mm. Its object-plane sampling is 0.1 mm per pixel. Using a projected tread speed of 10 m/s and allowing 1 pixel of blur:

These are application assumptions, not specifications for the example cameras. Confirm the usable pixel width, minimum exposure, trigger timing and achievable image quality before choosing hardware.

A higher lane speed, tighter sampling or stricter blur allowance shortens the available exposure further. If adequate illumination is impractical, reduce the permitted speed or change the imaging geometry rather than accepting training images that conceal the relevant features.

Light the grooves without making reflections the classifier

Black rubber needs controlled illumination, but maximum brightness is not the only objective. The image must preserve groove edges and tread-block structure without saturating reflective areas.

Start with bar lights positioned obliquely to the visible tread band. Change angle and distance while holding camera settings fixed. Opposing directions can reduce one-sided shadows, although excessive fill can flatten the relief cues you need.

The LS-BAR-1047 is a bar-light candidate for geometry trials. Its suitability depends on coverage, irradiance and mounting distance, which must be established with samples.

Use the lighting finder to compare relevant formats after establishing the geometry. Test both dry and wet tread appearances: reflections can obscure grooves even when the average image brightness looks acceptable. The scratch-lighting guide provides useful background on how illumination direction changes surface-feature contrast, though tire classification requires its own trials.

Short exposures may require synchronized pulses. Review light controllers and the controller finder for the required output type and documented timing behavior. The LS-CTL-0070 lists a strobe feature, but that alone does not establish compatibility with the pulse duration calculated above.

Confirm the light’s permitted pulse conditions, controller output capability, trigger electrical levels and camera exposure timing as a complete chain. Follow documented operating limits; do not infer an overdrive rating from average power.

The strobe timing and duty-cycle guide covers that selection process. If ambient light contributes materially during the exposure, a short pulse alone will not eliminate motion blur.

Capture tire events, not an uncontrolled stream

Associate images with a specific vehicle passage and tire position. The capture sequence should identify when useful tread enters the inspection zone and collect enough valid views to support a tire-level decision.

Frame rate alone does not guarantee coverage. As a simplified timing example, a 200 mm capture window moving past the camera at 5 m/s lasts:

The wheel’s rotation changes which tread patch is visible during that interval. Verify the complete sequence using representative passages rather than assuming successive frames show equivalent material.

Apply quality checks before classification. Useful checks include:

• Enough visible tread area, rather than mostly sidewall or background.
• Adequate sharpness within the tread region.
• Acceptable saturation and shadow coverage.
• A crop scale and viewing angle inside the validated range.
• Complete association with the correct tire position and passage.

Set quality limits using validation data. Do not invent a universal sharpness threshold that changes meaning with image size or processing.

For repeated captures, the streaming settings guide explains how configuration changes can interrupt acquisition. Freeze the validated exposure, gain, preprocessing and trigger configuration for production unless an adaptive mode has also been validated.

Build training data around real variation

Collect examples across tire families, wear conditions, vehicle geometries, tread contamination, illumination variation and the full permitted speed range. Include difficult negative examples that resemble the intended winter class. Include all-season and unusual patterns according to the labeling policy rather than forcing them into an unsuitable category.

Record capture conditions and label provenance. Retain enough information to distinguish a genuinely new tire from another frame of an existing one. Where operationally possible, retain tread crops instead of unrelated identifying imagery, and apply an appropriate retention policy.

Split data by physical tire and related vehicle passages before training. Adjacent frames from one passage must not appear in both training and test sets. Otherwise, the reported result can mostly measure recognition of nearly identical images.

Also separate tire-pattern families where practical. A model that succeeds only on patterns seen during training may fail on a newly introduced design. Report known-pattern and unseen-pattern performance separately when that distinction matters operationally.

Use training data for fitting, validation data for model and threshold selection, and an untouched test set for the final estimate. After repeatedly inspecting and tuning against a test set, it is no longer an independent final test.

Image augmentation can help represent modest changes in brightness, scale or orientation. It cannot replace missing tire categories, missing viewing geometries or real contamination examples. Keep synthetic transformations physically plausible for the station.

Validate the false-accept rate at the tire level

Define a false accept explicitly: a non-winter tire receives a supported winter classification. Evaluate that event after the entire pipeline, including image-quality rejection, crop aggregation and confidence thresholds.

Report at least the following:

A model score is not automatically a probability of correctness. A threshold of 0.98 does not establish a 98% reliable decision. Choose thresholds from held-out results and the permitted error budget, then evaluate the resulting manual-inspection workload.

Zero observed errors does not establish zero risk. With zero false accepts among N independent non-winter test decisions, an approximate one-sided 95% upper confidence bound is:

This approximation assumes representative, independent trials. Correlated tires from the same vehicle or repeated views of one tire cannot simply be counted as independent evidence. Account for clustering when estimating uncertainty.

Break down results by challenging conditions instead of reporting only an overall average. An acceptable aggregate result can hide poor performance on wet tread, unfamiliar patterns or the fastest permitted passages.

Route uncertainty and verify the whole station

Combine multiple crops with a documented tire-level rule. Agreement between neighboring frames is useful, but those frames are correlated. Do not multiply their confidence scores as though each were an independent observation.

If valid views disagree, required tire positions are missing, or the input lies outside the validated operating range, produce an uncertain result. Equipment faults, stale results and communication loss must not silently reuse a previous vehicle’s classification.

Begin commissioning in shadow mode: record classifications while the established inspection process remains responsible for decisions. Compare disagreements against independently reviewed labels. Change one acquisition or model variable at a time so improvements remain attributable.

Before ordering and acceptance, check:

• Visibility: representative vehicles expose sufficient tread at every required position.
• Sampling: the smallest useful features remain resolved after cropping and preprocessing.
• Motion: maximum permitted speed still produces usable tread edges.
• Lighting: reflections and shadows remain within the validated image-quality limits.
• Compatibility: camera, light and controller documentation support the required timing and electrical interface.
• Validation: independent test data support the selected false-accept limit and referral workload.
• Fallback: uncertain, incomplete and failed captures reach a defined manual inspection procedure.

Tread appearance classification is the wrong answer when the actual requirement is groove-depth measurement, damage assessment or legal approval of a winter tire. Those tasks need their own evidence, methods and acceptance criteria. A classifier must not substitute for a required measurement or safety-related inspection.

Start the camera configuration process with the field width, smallest feature, lane speed and required tire positions. For a capture-chain review, send those requirements and representative tread samples through application support before committing to hardware.