| --- |
| language: |
| - en |
| pipeline_tag: image-classification |
| library_name: pytorch |
| tags: |
| - industrial-vision |
| - factoryvision-ai |
| - casting-defect-detection |
| - transfer-learning |
| - mobilenetv3 |
| - education |
| --- |
| <div align="center"> |
| <img src="hero.svg" alt="FactoryVision AI MobileNetV3-Small casting inspection" width="100%"> |
| <p><strong>Educational prototype</strong> | Transfer learning and binary image classification | Human review required</p> |
| </div> |
|
|
| <h2>Model overview</h2> |
| <table> |
| <thead><tr><th>Model</th><th>Task</th><th>Framework</th><th>Weights</th></tr></thead> |
| <tbody><tr><td>FactoryVision AI MobileNetV3-Small</td><td>Casting defect classification</td><td>PyTorch / torchvision</td><td><code>model.pth</code></td></tr></tbody> |
| </table> |
|
|
| <h2>Held-out test results</h2> |
| <table> |
| <thead><tr><th>Metric</th><th>Reported result</th></tr></thead> |
| <tbody> |
| <tr><td>Accuracy</td><td><strong>0.9944</strong></td></tr> |
| <tr><td>Macro precision</td><td><strong>0.9932</strong></td></tr> |
| <tr><td>Macro recall</td><td><strong>0.9948</strong></td></tr> |
| <tr><td>Macro F1</td><td><strong>0.9940</strong></td></tr> |
| <tr><td>Mean inference latency</td><td><strong>0.395 ms</strong></td></tr> |
| </tbody> |
| </table> |
| <p><strong>Confusion matrix</strong> — rows are true labels, columns are predictions; class order: <code>def_front</code>, <code>ok_front</code>.</p> |
| <pre><code>[[450, 3], |
| [ 1, 261]]</code></pre> |
| <p>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.</p> |
| |
| <h2>Purpose and limitations</h2> |
| <p>Educational transfer-learning binary casting inspection (<code>def_front</code>, <code>ok_front</code>). <strong>Human review is required.</strong> This is not a validated or certified automated quality acceptance system.</p> |
|
|
| <h2>Model details</h2> |
| <ul> |
| <li>torchvision MobileNetV3-Small with a pretrained backbone and replaced classification head, fine-tuned on casting images</li> |
| <li>Packaged weights: <code>model.pth</code> (archive path <code>checkpoints/casting_mobilenet_v3_small.pth</code>)</li> |
| <li>Validation-selected best casting architecture in the supplied manifest: <code>mobilenet_v3_small</code></li> |
| <li>Class order: <code>labels.json</code>; for architecture and preprocessing see <code>src/quality.py</code> in the source repository</li> |
| </ul> |
| |
| <h2>Training data</h2> |
| <p>Real-Life Industrial Dataset of Casting Product by Ravirajsinh Dabhi (<a href="https://www.kaggle.com/datasets/ravirajsinh45/real-life-industrial-dataset-of-casting-product">Kaggle dataset</a>). 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.</p> |
| |
| <h2>Inference and dependencies</h2> |
| <p>Use the model-loader and preprocessors from <code>src/quality.py</code> in the <a href="https://github.com/Theivanth/Factory-Vision-AI">original GitHub project</a>. The checkpoint is a PyTorch artifact, <strong>not</strong> a drop-in Hugging Face Transformers model. Avoid loading untrusted serialized checkpoints.</p> |
| |
| <h2>Licensing and distribution</h2> |
| <p>The source code's MIT licence is distinct from trained-weight rights. <strong>Keep private until the casting dataset conditions and weight publication rights are clarified.</strong> No public checkpoint licence is asserted.</p> |
| |
| <hr> |
| <h2>Credits and attribution</h2> |
| <table> |
| <thead><tr><th>Role</th><th>Contributor</th></tr></thead> |
| <tbody> |
| <tr><td>Original Developer & Model Author</td><td>Theivanth Rajamohan</td></tr> |
| <tr><td>Technical Reviewer, Code Contributor & Publication Support</td><td><a href="https://huggingface.co/Swapinvidya">Swapin Vidya</a></td></tr> |
| <tr><td>Project Contributor</td><td><a href="https://huggingface.co/rajahshan">Rajah Shanmugasundaram</a></td></tr> |
| <tr><td>Publishing Organisation</td><td>TechSphere Consultancy Services Private Limited</td></tr> |
| </tbody> |
| </table> |
| <p><strong>Project provenance:</strong> FactoryVision AI was originally developed by Theivanth Rajamohan. <a href="https://huggingface.co/Swapinvidya">Swapin Vidya</a> 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.</p> |
| <ul> |
| <li><a href="https://huggingface.co/TechsphereAI">TechSphereAI on Hugging Face</a></li> |
| <li><a href="https://techsphereconsultancy.com/">Official website</a></li> |
| <li><a href="https://github.com/Theivanth/Factory-Vision-AI">Original GitHub project</a></li> |
| <li><a href="https://github.com/Theivanth/Factory-Vision-AI/pull/1">Swapin Vidya's merged contribution</a></li> |
| </ul> |
| |
