AI Can Now Identify Fabric From a Photo, Cutting Clothing Costs

When a clothing brand sources fabric, someone still has to sort incoming swatches by hand and type in what kind of weave it is — jersey, twill, satin. It's slow, repetitive, and a real bottleneck in getting clothes made. Researchers from Hong Kong and the UK trained AI to do that recognition automatically from an image, which sounds simple until you try to make it work outside the lab.

Their first finding is a warning that applies far beyond textiles. Earlier AI tests looked nearly perfect, but the team discovered the same fabric photos had sneaked into both the practice and test sets — like studying with the exact questions that appear on the exam. Once they rebuilt an honest test, accuracy dropped to 58%, showing what the problem really looks like.

The bigger hurdle wasn't confusion between similar fabrics, but the gap between how fabric looks in a catalogue versus in a real shipment. A simple fix — telling the AI to focus on the center of the fabric image rather than the edges — added 13.5 percentage points of accuracy and made the system's confidence scores trustworthy. Meanwhile, fancier approaches like hierarchical classification, text recognition, and zero-shot vision-language models all failed to help. The researchers reported those dead ends openly, which is good practice.

Finally, they tackled the cost question. Mixing up a satin for a twill matters more than a small slip within the same family, so they weighted mistakes accordingly. The system auto-labels swatches it's sure about and sends only the uncertain minority to a human. For fashion brands and textile suppliers, that means fewer people doing tedious typing, faster sampling rounds, and fewer costly mistakes — though human oversight still remains part of the process.