Learning

Synthetic steel surface defect classification

Classify six synthetic surface patterns

Train an image classifier on 600 steel-like images generated by Artelnics. The six pattern labels are crazing, inclusion, patches, pitted surface, rolled-in scale and scratches.

600Source records

40,000Encoded input values

120Testing-role records

6Model outputs

Review the current model Download model and data (ZIP)

Industrial challenge
Data set
Model
Training
Model selection
Testing
Deployment
Limitations

1. Industrial challenge

Manufacturing teams can use a reproducible image workflow to study class confusion and preprocessing. This example demonstrates image classification using generated textures; it does not measure defect detection on inspected steel.

Six pattern classes

Inspect errors between synthetic categories.

Repeatable generation

Retain the seed-based source generator.

Transfer evaluation

Use real acquisition data for industrial validation.

Quality engineeringComputer visionManufacturing

Synthetic image-classification demonstration, not validated industrial inspection.

2. Data set

The dataset contains 100 generated 200 × 200 grayscale images per class. The generation script and CC BY 4.0 notice are included in the project package. These are synthetic patterns, not the former NEU surface-defect photographs.

Source: Artelnics Synthetic Steel Surface Defects. Dataset license: CC-BY-4.0. The downloadable ZIP includes the adapted data and attribution notices.

Dataset measure	Saved value
Analysis unit	generated image
Records	600
Raw variables	2
Encoded model inputs	40,000
Model outputs	6
Training roles	360
Validation / selection roles	120
Testing roles	120
Unused roles	0

Field	Role	Type	Categories
image	Input	Numeric
Class	Target	Categorical	crazing, inclusion, patches, pitted_surface, rolled-in_scale, scratches

Target class distribution pie chart. Native Neural Designer report for this project.

The saved project assigns rows to training, validation and testing as shown above. This is internal record-level evaluation; it does not demonstrate separation by subject, device, site or acquisition batch.

3. Model

The model has 40000 encoded inputs and 6 outputs. No architecture-selection experiment is recorded in this project. The diagram shows the topology used by the saved model.

Layer	Input shape	Output shape	Activation
Scaling	200 × 200 × 1	200 × 200 × 1
Convolutional	200 × 200 × 1	200 × 200 × 8	ReLU
Pooling	200 × 200 × 8	100 × 100 × 8
Flatten	100 × 100 × 8	80000
Dense	80000	128	ReLU
Dense	128	6	Softmax

Output values are uncalibrated model scores. Use the output encodings and decision rule documented with this project; do not assume independent sigmoid scores sum to one.

Synthetic steel surface defect classification — initial network architecture — Topology of the saved current model; no architecture selection is recorded.

4. Training strategy

The saved training configuration uses Adam with CrossEntropy.

Adaptive moment estimation results

Measure	Value
Epochs number	200
Elapsed time	00:00:06
Stopping criterion	Maximum epochs number
Training error	0.035
Validation error	0.657

Adaptive moment estimation error history. Native Neural Designer report for this project.

5. Model selection

No model selection experiment is recorded for this version. The validation subset guides fitting where a training report is present; it is distinct from the held-out test rows.

6. Testing analysis

The figures and tables below refer to the current project’s saved testing analysis. The subset uses testing role 2; it contains 120 source records.

The multiclass decision uses the largest output score. The confusion matrix gives class prevalence and per-class errors; binary sensitivity, specificity, precision and a single ROC AUC are not interchangeable with this multiclass accuracy.

Confusion table

Measure	Predicted crazing	Predicted inclusion	Predicted patches	Predicted pitted_surface	Predicted rolled-in_scale	Predicted scratches	Total
Actual crazing	21 (17.5%)	1 (0.8%)	0	2 (1.7%)	0	0	24 (20.0%)
Actual inclusion	0	10 (8.3%)	12 (10.0%)	0	0	0	22 (18.3%)
Actual patches	0	0	16 (13.3%)	0	0	0	16 (13.3%)
Actual pitted_surface	0	7 (5.8%)	0	10 (8.3%)	0	0	17 (14.2%)
Actual rolled-in_scale	0	0	0	0	22 (18.3%)	0	22 (18.3%)
Actual scratches	0	0	0	0	0	19 (15.8%)	19 (15.8%)
Total	21 (17.5%)	18 (15.0%)	28 (23.3%)	12 (10.0%)	22 (18.3%)	19 (15.8%)	120 (100.0%)

Test measure	Value
Testing records	120
Accuracy	81.67%
Largest-class baseline	20.00%

7. Model deployment

Open the downloaded project in Neural Designer, inspect the dataset roles and preprocessing, then review the saved task report. Use the same input schema and category order when calculating outputs. The ZIP contains the exact current .nd, its source data and the applicable dataset notices.

Workflow: source measurements → schema and availability checks → model output → domain review. Keep model versions, validation evidence and incoming-data monitoring together.

Download model, data and licenses (ZIP)Download Neural Designer

8. Scope and limitations

A random split of images from the same generator measures within-generator performance. Validate with independently labelled real surfaces, different acquisition conditions and lot-separated testing before production inspection. Similar appearance does not establish the physical cause of a defect.

References

Artelnics Synthetic Steel Surface Defects. Artelnics. Not assigned
Dataset terms: Creative Commons Attribution 4.0 International. Full attribution and transformations are included in LICENSES/DATASET-LICENSE.txt.
Current Neural Designer project and saved task report, snapshot 6 October 2026.