Home/ Tool index/ The Squint Test
Launch ready Session 2 · Images Interactive

The Squint Test (feature extraction)

Change detail level and representation mode to see what survives when an image becomes numbers. The tool makes recognition thresholds visible — you can watch the moment an image stops being identifiable and ask which cue carried it longest.

tools/feature-extraction-pixel-resolution/

Live preview · launch for the interactive version

§ A · What it makes visible

what the screen literally shows
Fig. 01

Images as numbers

Before a model can see anything, an image is a grid of numbers. Feature Extraction makes that grid inspectable — you can see exactly what the model is working with at any level of detail.

Fig. 02

Recognition thresholds

Reduce detail until the image stops being recognizable. The last surviving cue tells you what the recognition system is relying on — often not what you would expect a human to use.

Fig. 03

Which cues carry recognition

Color, outline, texture, context, and spatial arrangement each carry different amounts of recognition information. This tool lets you test the hierarchy directly by watching what survives.

§ B · How to investigate it

run it like an experiment, not a toy

Reduce detail one step at a time. Describe what you see at each level — not what you expect to be there.

01 · Predict

Before reducing detail

Write down which cue you think will carry recognition longest: color, outline, texture, or spatial context. Then test your prediction.

prediction: “outline will survive longest”
02 · Reduce systematically

One level at a time

Drop one level of detail at a time. At what level does recognition become uncertain? At what level does it fail completely?

level 5 → uncertain · level 3 → unrecognizable
03 · Compare image types

Does the same cue survive?

Run the same reduction on two different images. Does the same feature type survive longest, or does it depend on the subject?

face vs. object · same reduction, different survival
04 · Name the surviving signal

Not “it’s blurry”

Name the specific feature that still carries identity after everything else is gone. That feature is what the recognition system is relying on.

“color distribution alone was enough for this image”

§ C · Debrief questions

after the investigation
Which visual cue survived longest as resolution dropped?
What disappeared first, and what lasted longer than you expected?
What does the model rely on that a human reader might not notice?
Where could a surviving-but-wrong signal cause a misidentification?