Learning & Labeling
WAION Vision AI confirms feasibility on a small initial dataset, then completes the inspection functions a process needs step by step, repeating AI labeling and on-site verification.
Step-by-step learning from about ten images
When the inspection target and imaging conditions are reasonably clear, we do not gather a large dataset first. A small one confirms feasibility, and the training scope then grows to match production conditions.
01
About ten representative images
With a clear target and imaging conditions, a PoC starts from roughly ten representative images.
02
Data built around the key variables
The dataset is organized by the variables that matter: lighting, product position, normal and abnormal states, and occlusion during work.
03
Missed and false detections analyzed
Missed and false detections from the first model are analyzed, and only the cases needed are added, step by step.
04
Scope grown to production conditions
Once feasibility is confirmed, the training scope widens to match production conditions.
In-house AI labeling and review
An AI labeling tool built in-house generates pre-labels, and operators correct only what needs it. Label quality is managed through review history and data versions.
01
Pre-labels generated
An existing model proposes the position and class of products and parts as pre-labels.
02
Checked and corrected by operators
Operators confirm the proposals and correct only what needs it, cutting repetitive labeling time.
03
Review history and data versions
Label quality is managed against review history and data versions.
Reviewed data
Split into training and independent evaluation sets
- Training
- Used to train the model.
- Independent evaluation
- Data the model never trained on, used to check it.Checked in independent evaluationMissed detectionsFalse detectionsVerdict accuracy
Contact
Leave us a message and our team will get back to you shortly.
See OK / NG / NR verdicts on your actual products. From planning to installation and setup, we're with you on-site.