Negative samples reduce false positives: train with background images

A detector that fires on empty fixtures, shadows or the neighbouring conveyor does more damage than one that occasionally misses a part: every phantom box is a false reject, a stopped line or an operator who stops trusting the screen. The fix lives in the training set, not in the confidence threshold. Negative samples are background frames with no parts and no boxes; added at the right fraction they teach the model what an empty scene looks like. The working numbers are 0-10% of the training set, starting near 5%, and this guide pairs them with a capture recipe and the arithmetic that proves the false-positive drop before you deploy.

What a negative sample is

A negative sample is a frame your production camera took of a scene with zero target objects, stored with zero bounding boxes. Same camera, same lens, same lights, same mounting. A picture of a similar cell pulled from the web teaches the wrong textures and helps nothing.

Three kinds of frame are not negatives:

• A frame where a part is visible but unlabelled. The trainer reads the missing box as a miss and trains itself to ignore your part. Label it or discard it.
• A frame where the part sits outside the inspection region but on the sensor. That is a normal labelled frame; the sensor still sees the object.
• A frame from a different camera or lighting rig. Its background distribution is not the one your deployed model will meet.

The labels stay empty on purpose. Why the empty frame needs no boxes, and how the trainer reads it, is laid out step by step in Background images in object detection training: why no labels.

Why empty frames cut false positives

A single-stage detector scores every image location as background or object. Train it only on frames full of parts and the ambiguous textures — a shadow with a soft edge, a fixture that catches the light — lean towards a box, because the training distribution says every frame holds something.

Empty frames cast the other vote. Each one contributes thousands of location decisions where the correct answer is "nothing here", and the confidence of phantom boxes falls.

The evidence has been around for years. A 2020 study scrambled image backgrounds and watched accuracy fall by 6-14% on ImageNet benchmarks and 13-22% on a nine-class set: the models were reading the background, not only the object. A 2023 clinical detection study reached the applied version of the same conclusion — false positives survived every tuning pass until negative samples joined the set. The reference COCO dataset ships about 1,000 deliberately unlabelled background images, roughly 1% of its training set, for exactly this reason.

What the non-object area inside your labelled frames already teaches, and how to keep that distribution honest, is the subject of Object detection background distribution for inspection.

How many background images to add

Published training guidance for YOLO-style detectors puts the working range at 0-10% of the training set; practitioner runs land at 1-10% with 5% as the common starting point. The table turns that into a decision:

Which frames to capture as negatives

Shoot the negatives in the cell, across the states the camera actually sees:

• Empty fixture and empty pallet positions, including the missing-part state your reject logic depends on.
• The neighbouring machine, conveyor edge, guarding and cables that sit in view.
• Shadows and reflections as hall and daylight move through the shift.
• Gloves, sleeves and tools at the frame edge, if operators work near the field of view.
• Any state that produced a phantom box in your logs. That frame is the highest-value negative you own.

Cover the light range, not one hour of it; how to hold ambient light steady, and what to do when you cannot, is in Ambient light variation in pick and place vision: control it. Around a hundred frames per scene state is enough for training; the hold-out set that measures the result needs its own sizing, below.

Empty labels, not a background class

Give every background frame an empty label file — present in the labels folder, with no boxes inside. Some trainer versions skip images whose label file is missing, which silently removes your negatives from the run; an empty file is read as a zero-object frame on every version.

You will meet the advice to add an explicit background class instead. It can work, but it invites boxes drawn around empty regions and couples your class list to a catch-all. Empty frames reach the same goal and leave the class list clean.

Measure the drop before you deploy

Two ratios carry the whole decision:

Both are dimensionless; report them as percentages. Count frames, not boxes: a frame with three phantom boxes is one bad frame. Freeze the empty hold-out before the first run and never train on it, so every model is scored on the same frames.

Worked example (numbers are illustrative; the method is what you copy):

• Training set: 2,400 labelled part frames. Add 120 background frames: 120 / 2,520 = 4.76%, call it 5%.
• Empty hold-out: 400 frames captured at the same mount and lighting, split across shifts.
• Baseline model, positives only: 18 of 400 hold-out frames show at least one phantom box above the deployment threshold. 18 / 400 = 4.5%.
• Retrained model: 6 of 400. 6 / 400 = 1.5%.
• Recall check on 600 part frames: 576 / 600 = 0.960 before, 573 / 600 = 0.955 after, a cost of half a point.

Shift the same numbers to line impact: at 400 empty-check positions per shift, 4.5% means 18 phantom rejects per shift and 1.5% means 6. At one minute to acknowledge and clear each stop, the retrain hands back 12 minutes of line time per shift.

Decision: 4.5% to 1.5% at half a point of recall clears the bar — ship the retrained model. If the rate had stayed above 2%, the next move is a background pass at 8-10%, not a threshold change; thresholds come last, as Reduce object detection false positives in a new scene works through. Score every run with the same matrix on the same hold-out — Interpret an object detection background confusion matrix shows which cell is which.

Where the retrained model runs

Retraining is a compute job, not a line job: run it off-line and leave the production controller in service. A dedicated training machine such as the AI training server below is built for this loop — dataset in, weights out, no downtime on the cell.

Once the false-positive target is met, the weights ship to the hardware that watches the cell: an AI inference terminal mounted beside the cameras, or a smart camera that runs the detector on the device. The camera finder filters that line by resolution, interface and shutter if you are specifying hardware for a new cell.

On-device inspection suits small single-class detectors — presence, position, one-part checks — which is exactly the shape of model this article trains. The two cameras below are common starting points, and the vision controllers take over when several cameras or a PLC handshake join the loop.

When background images are not the fix

Two root causes do not answer to negatives:

• The false positive looks like the part because your positives never showed that variation — shiny coatings, bent flanges, parts at the frame edge. The cure is more positive variety and harder examples, the territory of One class fails in object detection training: fix rare small classes. Empty frames cannot teach your part its own appearance.
• There are no boxes to draw at all — surface stains, texture drift, whole-image anomalies. That is an appearance problem, and an anomaly-detection setup fits it better; Anomaly detection without ground truth: test and heatmap setup shows the path.

If the model misfires only in a scene it has never seen, that is domain shift first; negatives from the old cell will not save it.

Before you retrain: the checklist

• Capture at least 200 empty-scene frames as a frozen hold-out. At 200 frames one extra phantom box moves the rate by 0.5 points, small enough to read a trend.
• Log every phantom box the current model fires, and shoot that exact state as a training negative.
• Add backgrounds at 5% of the training set, all from your cell, all with empty label files.
• Keep epochs, image size and augmentation identical to the baseline run, so the only changed variable is the negatives.
• Re-score the hold-out and report the false-positive frame rate and recall side by side.
• Move to 8-10% only if the rate is still above target; hold at 3% and add positive variety if recall slips.

More systems-side reading, including the deployment and data-handling guides this one leans on, is collected on the systems topic page. To price the hardware path — camera, inference terminal, trigger and lights as one order — the build-a-system configurator pairs the parts, and the quote form returns a bill of materials for review.