Glass is semi-transparent, nonwoven fibers are chaotic, and defect samples are both scarce and scattered. DaoAI pairs few-shot learning with good-only modeling so the system goes live reliably even without ample defect data.
This materials plant produces both glass and nonwoven—two surfaces with vastly different characteristics. Glass is semi-transparent and prone to background and glare interference, with bubbles, stones and scratches showing weak contrast; nonwoven has naturally chaotic fiber distribution, a classic weak-texture surface where defects are hard to separate from background. Their shared challenge: low defect occurrence means trainable defect samples are both scarce and highly variable.
Supervised deep learning needs abundant labeled samples, but with defects so rare the data simply cannot be assembled, and projects often stall for lack of it. Manual visual inspection, hampered by semi-transparency and weak texture, suffers fatigue-driven misses over long shifts and poor consistency.
The DaoAI Few-shot + APDT Anomaly Detection Solution
DaoAI combined few-shot learning with APDT good-only anomaly detection: when defect samples are scarce, it first builds a normal baseline from ample good images, then fine-tunes the criterion with a tiny number of defect samples. For semi-transparent glass and weak-texture nonwoven, the model separates defect signals under low-contrast, weak-texture conditions, going live fast without large-scale labeling and reaching 94%+ recall.
- Goes live with few shots—no need to assemble a large labeled dataset for rare defects
- Built on a good-product normal baseline, fine-tuned with minimal defect samples, also catching unseen defects
- Optimized for semi-transparent glass and weak-texture nonwoven, effectively separating low-contrast defects from background
- Recall above 94% on glass bubbles/stones/scratches and nonwoven holes/foreign matter
The scarcer the defect samples, the greater the value of learning good only—build the normal baseline first, then top up the criterion with a few shots.
After go-live, the plant brought glass and nonwoven inspection online quickly despite scarce defect samples, with recall stable above 94%. Hidden defects on semi-transparent and weak-texture surfaces were missed far less often, inspection moved from experience-based visual checks to standardized online judgment, new defect types could be added quickly from a few samples, and both consistency and maintainability improved together.
FAQ
What difficulties are faced in the surface defect detection of glass and non-woven fabrics?
Glass is translucent, easily affected by background and reflection, with weak defect contrast. Non-woven fabrics have disordered fibers, a typical weak-texture surface, making it difficult to separate defects from the background. Also, defect samples are scarce and varied.
How does DaoAI solve the problem of scarce defect samples?
DaoAI combines few-shot learning with APDT positive sample anomaly detection. It first establishes a normal baseline with sufficient good product images, then fine-tunes the criteria with a very small number of defect samples, enabling quick deployment with few samples.
What are the effects after the implementation of DaoAI's detection solution?
The factory quickly launched the detection of glass and non-woven fabrics with scarce defect samples. The defect detection rate reached over 94%. Hidden defect misses decreased, quality inspection became standardized online judgment, and consistency and maintainability improved.