AI Can't Yet Count Threads in Fabric Photos Reliably
If you buy fabric, this AI limitation could affect quality and price.
When you buy a shirt or curtains, the fabric's thread count is a key quality indicator. Inspectors typically count threads manually using a magnifying glass called a pick glass, which is slow and tedious. A smartphone photo through that glass could speed things up and create a record. But can AI accurately count threads from such photos? A new study created a dataset of 76 factory scenes and tested various AI models. The results were disappointing: no AI method came within the required 5% tolerance. The best models were off by about 25%, meaning they'd often be wrong by a quarter of the actual count. Even a non-expert human did better than some AI, though still not within tolerance.
The study also found that a training-free method called PG-Lattice was most consistent across repeated photos, but its errors were often off by a factor of two. Even the human reference counts varied by a median of 9% across re-shoots, suggesting that one photo isn't enough to meet the 5% tolerance. This means that reliability across multiple photos is what matters for image quality, not just a single shot. Generative enhancement, which makes photos look better, didn't improve counting accuracy. And generic image quality scores didn't predict which method would be more accurate.
For the textile industry, this means AI isn't ready to replace human inspectors for thread counting. However, the study shows that a small AI model tuned on synthetic images can match the best frontier models, which is promising for future development. The key takeaway is that task-based image quality—whether the image serves its purpose—is more important than perceptual quality. Until AI can reliably count threads within tolerance, human expertise remains essential. This research highlights the challenges of applying AI to precise measurement tasks and the need for specialized approaches.
- AI models tested could only count fabric threads within 25% accuracy, far from the 5% needed for quality control.
- A small AI model tuned on fake images performed as well as the best general AI, showing potential for specialized training.
- Even human counts varied by 9% across repeated photos, so one photo isn't reliable enough for precise measurement.
Why It Matters
For now, AI can't replace human inspectors for fabric quality, so expect manual checks to continue.