Educational prototype | Transfer learning and binary image classification | Human review required
Model overview
| Model | Task | Framework | Weights |
|---|---|---|---|
| FactoryVision AI MobileNetV3-Small | Casting defect classification | PyTorch / torchvision | model.pth |
Held-out test results
| Metric | Reported result |
|---|---|
| Accuracy | 0.9944 |
| Macro precision | 0.9932 |
| Macro recall | 0.9948 |
| Macro F1 | 0.9940 |
| Mean inference latency | 0.395 ms |
Confusion matrix — rows are true labels, columns are predictions; class order: def_front, ok_front.
[[450, 3],
[ 1, 261]]Metrics are from the saved training output, not independently reproduced. Latency is environment-specific and is not a deployed CPU benchmark. The test split did not determine architecture selection; the selection metric was validation macro F1.
Purpose and limitations
Educational transfer-learning binary casting inspection (def_front, ok_front). Human review is required. This is not a validated or certified automated quality acceptance system.
Model details
- torchvision MobileNetV3-Small with a pretrained backbone and replaced classification head, fine-tuned on casting images
- Packaged weights:
model.pth(archive pathcheckpoints/casting_mobilenet_v3_small.pth) - Validation-selected best casting architecture in the supplied manifest:
mobilenet_v3_small - Class order:
labels.json; for architecture and preprocessing seesrc/quality.pyin the source repository
Training data
Real-Life Industrial Dataset of Casting Product by Ravirajsinh Dabhi (Kaggle dataset). The source README identifies CC BY-NC-ND 4.0 dataset restrictions, including non-commercial and no-derivatives clauses. Rights to redistribute trained model weights remain to be confirmed; do not redistribute source images.
Inference and dependencies
Use the model-loader and preprocessors from src/quality.py in the original GitHub project. The checkpoint is a PyTorch artifact, not a drop-in Hugging Face Transformers model. Avoid loading untrusted serialized checkpoints.
Licensing and distribution
The source code's MIT licence is distinct from trained-weight rights. Keep private until the casting dataset conditions and weight publication rights are clarified. No public checkpoint licence is asserted.
Credits and attribution
| Role | Contributor |
|---|---|
| Original Developer & Model Author | Theivanth Rajamohan |
| Technical Reviewer, Code Contributor & Publication Support | Swapin Vidya |
| Project Contributor | Rajah Shanmugasundaram |
| Publishing Organisation | TechSphere Consultancy Services Private Limited |
Project provenance: FactoryVision AI was originally developed by Theivanth Rajamohan. Swapin Vidya contributed technical review, detection-metrics improvements, testing, model publication preparation and documentation. The models are being prepared for publication under TechSphere Consultancy Services Private Limited's Hugging Face namespace, with the original developer's permission.