Normal surfaces
240Retain textures, seams and reflections to define non-defect regions.
PROJECT FOCUS / MFG-001
Industrial vision · Case study
From surface variation to consistent defect boundaries
Use normal textures and defect samples to agree annotation rules for metal surfaces. Deliver region labels with indexes that support review.
← Back to case studiesDATA IN CONTEXT
600Images
4Label classes
1Pilot batch
THE BRIEF
Define what the data needs to support.
Organize images and labels for training and review in surface inspection. The priority is distinguishing defects from normal texture, rather than labeling every visual difference as a defect.
- Application
- Industrial vision
- Work involved
- Data preparation, annotation, quality review and delivery
- Task focus
- Surface defects · Region annotation · Rule alignment
SCENE COVERAGE
Break the scope into reviewable groups.
Scratch scenes
180Distinguish thin scratches from machining texture and annotate visible boundaries.
Dent scenes
120Use visible shading to determine extent without inferring hidden areas.
Stain scenes
60Separate discoloration, deposits and inconclusive appearances.
FROM CHALLENGE TO METHOD
Connect each decision to a record.
DECISION / 01
Glare mistaken for defects
01 / THE APPROACH
Include difficult samples in the pilot, record normal texture and uncertainty, and establish positive and negative examples.
02 / THE REVIEW TRAILClass dictionary, boundary references and difficult-sample log
DECISION / 02
Inconsistent defect boundaries
01 / THE APPROACH
Agree rules for visible regions, connected regions and small defects before batch annotation.
02 / THE REVIEW TRAILGuideline version and revision log
DECISION / 03
Similar images crossing dataset splits
01 / THE APPROACH
Group by component and capture batch, check near-duplicates and retain split rationale.
02 / THE REVIEW TRAILComponent index, deduplication list and split mapping
REVIEW & HANDOVER
Review the task. Verify the delivery.
Review focus
Class decisions
Check false defect labels on normal texture, missing defects and handling of uncertainty.
Region boundaries
Review contours, out-of-bounds regions and overlaps; do not equate region counts with image counts.
File integrity
Ensure every image has a status record and label coordinates match image dimensions.
Delivery contents
- 01
Image & component index
Link files, components, primary scenes and batches.
- 02
Region & state labels
Store normal states, defect regions and uncertainty fields.
- 03
Guidelines & review records
Include class definitions, issue lists and rule versions.
WHAT TO TAKE FORWARD
What matters for the next project
Agree the rules using normal and borderline samples before scaling defect annotation. Agree the data scope, labeling policy and review criteria together.
MORE PROJECT PERSPECTIVES
Explore other industry cases
YOUR NEXT DATA PROJECT
Bring your task into the discussion.
To frame a similar pilot, share: Task goals, product categories, capture conditions, defect definitions and shareable representative samples.