How the Laundry Symbol Scanner Works
The laundry symbol scanner runs two AI models on our server: one finds the symbols, one reads them. On 132 real tag photos it had never seen, it read 96% of symbols right and every symbol on 78% of tags.
Step 1: finding the symbols
The first model is a detector based on RF-DETR Nano, an open source model released by Roboflow under the Apache 2.0 licence. It looks at the photo scaled to 384 by 384 pixels and draws a box around everything that looks like a care symbol. It does not try to name them. Keeping its job simple makes it reliable: on the test photos it found 99.4% of the symbols.
Step 2: reading each symbol up close
A 384 pixel view of a whole tag leaves each symbol only a few pixels wide, too small to count dots. So the second model goes back to the full-resolution photo, cuts out each box, and reads it at 160 by 160 pixels. It names the symbol (42 kinds, plus "not a symbol" for logos and text), and counts the lines underneath. It reads each crop three times at slightly different framings and averages the answers, which steadies the dots and lines.
Step 3: writing the instruction
Many tags repeat the same row of symbols for different countries. Plain code keeps the most confident reading in each group (washing, bleach, drying, ironing, professional care) and writes them in that order, as one sentence.
What it learned from
The models learned from about 880 distinct photos of real care tags. We merged their many different label names into one set of 42 symbols, checked hundreds of doubtful labels by eye, and fixed or removed the wrong ones.
To show the models every symbol many times, we also drew all 42 symbols ourselves and printed them onto 8,000 synthetic tags with random fabric textures, fonts, ink colors, blur, glare and angles.
How we tested it
We tested on 172 phone photos of real tags from a separate source the models never saw during training. We checked every label by hand, corrected wrong ones and added about 300 symbols the original labels had skipped. We scored the 132 photos of real garment tags; the rest are photos of symbol charts on screens.
| Measure | Result |
|---|---|
| Symbols read exactly right, dots and lines included | 96.0% of 841 |
| Symbols the detector missed | 0.8% |
| Extra detections that were not symbols | 1 in 8 photos |
| Tags with every symbol right | 78.0% of 132 |
It was perfect on the most common symbols: do not bleach, do not tumble dry, do not dry clean, do not iron and hand wash. It was weakest on the wash tub at 30°C (87%), the P circle (91%) and iron on low (92%). Almost every mistake was one of the small marks: a line under the tub or circle seen or missed, or one dot read as two.
What it cannot do
- Read words. US tags sometimes write limits such as "no steam" or "short cycle" instead of drawing them.
- Read national symbols outside the international set, such as older Japanese and Korean symbols with written characters.
- Fix a photo that is too small or blurry to show the dots. Get close and keep the tag flat.
What happens to your photo
Your browser shrinks the photo to at most 2048 pixels on the long side, which keeps the dots sharp, and sends it to our server. Both models run there with onnxruntime, in about a second. The photo is held in memory only for that time and is never stored. See the privacy policy.
References
- ISO 3758:2023, Textiles: Care labelling code using symbols.
- ASTM D5489, Standard Guide for Care Symbols for Care Instructions on Textile Products.
- RF-DETR, real-time detection transformer by Roboflow, 2025. Apache 2.0, github.com/roboflow/rf-detr.
Have the tag in your hand? The scanner reads every symbol in seconds.
Open the scanner