Troubleshooting: False Positives in Automated Defect Detection

Why false positives appear in automated inspection

Automated defect detection systems flag non-defects when sensor input does not match the reference model. The machine sees a spot, a streak, or a color shift and classifies it as an error. Operators then stop the line or mark the fabric for rejection. This creates rework and delays.

False positives reduce trust in the system. If engineers see too many false alarms, they may disable the machine or ignore its output. The result is lower inspection accuracy and missed real defects.

The root causes usually fall into three groups: optical issues, environmental changes, and software settings.

Optical and sensor issues

The vision system relies on a clear optical path. Dust, lint, or oil on the lens blurs the image. The camera then sees noise that looks like a small defect.

Check the lens before changing software settings. A quick wipe with a dry, lint-free cloth removes most surface contamination. Do not use compressed air directly onto the sensor glass unless the housing is sealed. The air stream can push fine particles into the optical cavity.

Also check the focus. If the camera loses focus during long runs, the image softens. Soft edges create false edges in the detection logic. Thermal expansion can shift the focus position over several hours. A camera that was sharp at the start of a shift may be slightly out of focus by the end of the night run.

Another common issue is the lens coating. Cheap filters or protective glass can create glare under certain lighting angles. This glare shows up as a bright spot that the system flags as a defect. The angle changes as the fabric moves through the inspection head. What appears as a static bright spot in a still image may move across the field of view during a run.

Inspect the protective cover for scratches. A fine scratch on the glass acts like a tiny prism. It splits the light source into a faint line or arc. The detection algorithm may interpret this line as a yarn break or a thin defect.

Lighting and environmental factors

Lighting is the most common cause of false positives. If the light source flickers, the image brightness changes between frames. The system may interpret the brightness shift as a defect.

Check the light source type. Some LEDs emit inconsistent light if the driver is failing. The output may pulse or dim. This creates a moving artifact in the image. The artifact often follows a pattern related to the power supply frequency or the thermal state of the driver.

Ambient light also matters. If a door opens and daylight hits the camera, the exposure changes. The system may flag the whole frame as defective. In open factory floors, moving vehicles or people casting shadows can trigger similar responses.

Use a stable light source. Shield the camera area from ambient light. Keep the lighting consistent across the full fabric width. A light bar that is too short will create a gradient across the fabric. The edges of the fabric may appear darker than the center. The algorithm may flag this gradient as a color shift or a defect.

Check the color temperature of the light source. If the bulb has aged, the color temperature shifts. A white light that was 6500K may now read closer to 5500K. The color reference model will see this as a color deviation.

Software thresholds and model drift

The detection logic uses thresholds to decide what is a defect. If the threshold is too low, the system flags every tiny variation. This leads to false positives.

Model drift happens when the system learns the wrong pattern. If the operator trains the model on a single fabric type, it may flag normal weave variations in another type as defects.

Review the detection logic. Adjust the threshold to match the fabric. Use a reference sample of the same material to test the system. The reference sample should be from the same lot or batch as the production roll. Different lots can have slight color and texture variations. A reference from a previous lot may not match the current material.

Also check the model update process. If the system updates its model after a shift and the operator does not verify the update, the model may drift. Some systems update the model automatically based on operator corrections. If the operator marks a real defect as a false positive, the system learns to ignore that type of defect.

Check the confidence score settings. Some systems allow you to set a minimum confidence level for flagging a defect. If this level is too high, the system flags uncertain detections. If it is too low, it may miss real defects. Adjust this setting based on the criticality of the fabric.

Common symptoms and fixes

| Symptom | Likely cause | What to do |
| Symptom | Likely cause | What to do |
| Bright spots appearing in one area of the image | Dust on lens or light reflection | Clean lens and shield from ambient light |
| Whole frame flagged as defective | Lighting flicker or exposure shift | Check light driver and stabilize exposure |
| Repeated flags on the same fabric region | Camera focus loss or sensor misalignment | Re-focus camera and check mounting hardware |
| Flags appear only during high-speed runs | Motion blur or exposure time too long | Reduce roll speed or adjust exposure time |
| Random flags across the fabric | Software threshold too low | Increase threshold and verify with reference sample |
| Flags increase after model update | Model drift or incorrect training data | Review update log and retrain with current fabric |

Calibration and reference checks

Calibration is the first step when false positives increase. A vision system needs a known good reference to measure against.

Use a clean, defect-free section of the same fabric. Run the system on this section. If it flags the reference, the system is misaligned or the threshold is wrong.

Check the camera position. If the camera moves even a few millimeters, the field of view changes. The system may detect edges that were not there before. Vibration from nearby machinery can loosen mounting screws over time. Check the mounting bolts after any major maintenance or floor work.

Verify the exposure time. If the exposure is too short, the image is noisy. If it is too long, motion blur appears. Both create false edges. The optimal exposure time depends on the roll speed and the light intensity. A faster roll speed requires a shorter exposure time to avoid blur. A dimmer light source requires a longer exposure time to capture enough detail.

Check the white balance. If the white balance shifts, the color reference changes. The system may flag a neutral gray fabric as a color deviation. Set the white balance using a neutral gray card or a known white reference.

Operator feedback and alarm limits

Operators know the fabric better than the system. If they see a flag and the fabric is clean, that is a false positive.

Use operator feedback to tune the system. Collect flagged images. Review them with the operator. Mark which are real and which are false. The operator may see a pattern that the algorithm misses. They may notice that a specific type of flag appears when the fabric is wet or when the roll is near the end.

Set alarm limits based on this feedback. If a region flags too often, increase the threshold or create a special rule for that region. Some systems allow you to define exclusion zones. Use these zones for areas where the fabric naturally has more variation, such as the selvage or the center of the roll.

Do not let the system run unattended for long periods without review. A weekly check of flagged images keeps the system aligned with reality. Assign this task to a specific operator or technician. Keep a log of the checks. This log provides a history of the system performance.

Prevention tips

1. Clean the lens and sensor glass before each shift.
2. Shield the camera area from ambient light and open doors.
3. Check the light driver for flicker or dimming.
4. Verify exposure time and focus after any camera movement.
5. Recalibrate after changing fabric type or roll speed.
6. Review flagged images weekly and update thresholds based on operator feedback.
7. Keep a log of model updates and verify each one with a reference sample.
8. Train operators to recognize false positives and report them promptly.
9. Use stable mounting hardware for the camera and light source.
10. Test the system with a known good and known defective sample before production.

When to call a specialist

Some false positives point to hardware failure. If the lens is clean, the lighting is stable, and the calibration is correct, but the flags continue, the sensor may be failing.

Check the sensor for dead pixels or hot pixels. These appear as bright or dark spots in the image. A failing sensor creates artifacts that no amount of calibration will fix. Dead pixels are usually static. Hot pixels may change color or brightness over time due to heat.

Also check the data cable. A loose or damaged cable can cause intermittent signals. The image may drop out or show noise. This noise can trigger false flags. Inspect the cable for kinks, cuts, or loose connectors. Test the cable with a known good sensor if possible.

Check the power supply to the camera and light source. A failing power supply can cause voltage drops. This can affect the camera’s internal electronics and the light source’s output. Use a multimeter to check the voltage at the connector while the system is running.

If the issue persists after checking the optics, lighting, and software, contact the system vendor. They can test the sensor and replace failing components. Provide them with the specific symptoms, the fabric type, and the settings you have already tried.

Final checks before production

Before starting a new fabric run, perform a quick checklist.

1. Clean the lens and protective glass.
2. Verify the light source is stable and shielded.
3. Check the camera focus and mounting.
4. Run a reference sample through the system.
5. Review the last ten flagged images.
6. Confirm the threshold matches the current fabric.

This routine takes a few minutes. It prevents false positives before they start. It also builds operator trust in the system.

When the system flags only real defects, inspection accuracy improves. Operators stop the line less often. Rework drops. The line runs smoother.

Defect detection is not a set-and-forget process. It needs regular attention. The false positives you fix today will save time and money tomorrow.