In materials surface inspection, defect shapes are nearly infinite, and chasing them by collecting defect samples never keeps pace with the line. DaoAI shifts the problem from enumerating defects to modeling normality.
This materials plant produces functional surfaces with complex textures. The surface itself carries natural texture variation, while defects span dozens of forms—scratches, dents, contamination, local gloss loss—and new process batches keep producing anomalies never seen before. Supervised vision needs ample samples per defect class, but on a surface with nearly infinite defect shapes, positives and negatives can never be enumerated, and the model is obsolete the day it ships.
The team previously relied on manual sampling and rule-based algorithms. The textured background caused heavy under-detection of low-contrast defects, and rule thresholds failed repeatedly after batch changes, keeping rework and complaint costs high. The core contradiction: defects cannot be defined in advance, yet the system was asked to recognize every one of them.
The DaoAI APDT Good-only Anomaly Detection Solution
DaoAI deployed APDT good-only anomaly detection, training on good-product images alone and using the DaoAI World model to build a normal baseline of the texture. Any region that deviates from the baseline—whether or not it appeared in training—is flagged as an anomaly and localized at pixel level. Against the textured background, the model still separates the defect signal from natural texture under low-contrast, weak-texture conditions.
- Modeling needs good images only—no defect labeling—cutting deployment from weeks to days
- Image-level AUROC holds above 99%, with pixel heatmaps pinpointing defect locations
- Previously unseen defect types are caught without retraining; the normal baseline transfers to new batches
- Integrated with existing line cameras, recall on hidden defects like low-contrast scratches rose sharply
Shift from enumerating defects to modeling normality—no matter how infinite the defect shapes, none escape the normal baseline.
After go-live, the plant launched without collecting a single defect sample. Image-level AUROC stayed above 99%, batch changes required no re-labeling, complaints from under-detection dropped markedly, and inspection labor moved from full screening to anomaly review, raising overall consistency substantially.
FAQ
What are the problems with traditional supervised vision in material surface quality inspection?
In material surface quality inspection, the defect forms are almost infinite. Traditional supervised vision needs to collect sufficient samples for each type of defect. Positive and negative samples can never be exhausted, and the model becomes obsolete once launched. There are also problems such as missed detections and frequent failures of rule thresholds.
What are the advantages of DaoAI APDT positive sample anomaly detection solution?
DaoAI APDT solution only uses good product images for training, without the need for defect sample annotation, and the deployment cycle is short. The image-level AUROC is over 99%. It can locate new defects, connect with production line cameras, and improve the recall rate of hidden defects.
What are the effects after the implementation of the DaoAI solution?
The factory completed the launch without collecting any defect samples. The image-level AUROC has been consistently above 99%. There is no need for re-labeling when switching to new batches. Customer complaints due to missed detections have significantly decreased. The quality inspection manpower has been optimized, and the overall inspection consistency has been greatly improved.