A small set of normal images can support initial checks of image input, alignment, imaging variation and candidate methods. It is insufficient evidence of real-defect detection performance or complete acceptance coverage. Without real defective samples, keep missed-defect conclusions marked as unverified.
What normal samples can support
Check formats, sharpness, lighting, location variation and normal textures. Candidate-method false alarms can also be observed on the supplied images. Those observations apply to the supplied samples. They do not directly establish behavior for every conforming item, batch or production period.
Additional inputs needed
| Input | Purpose |
|---|---|
| Defect classes, dimensions and decision criteria | Define what must be detected |
| Real defective and borderline samples | Check defect coverage and misses |
| Different batches and normal variation | Check false alarms on conforming items |
| Unusable images and alignment-failure samples | Define review-required and failure states |
| Independent acceptance data and conditions | Produce acceptance records for the stated scope |
Does normal-only training remove the need for defect testing
Even when a candidate anomaly-detection method trains mainly on normal samples, real defects matching business definitions and independent data are still needed to verify detection performance. Synthetic defects can assist analysis. They do not automatically replace all real defects or field conditions.
Starting with incomplete data
Limit the product types, imaging conditions and pilot objectives first. Deliver a sample-gap list, candidate route and testing plan. Record checked content separately from unverified scope. Establish acceptance conclusions after the missing samples are supplied. An initial demonstration is not evidence of a complete inspection pass.
Related services: vision algorithms and model engineering and project enquiries.