| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - materials_engineering |
| - materials |
| - steel |
| --- |
| # UHCS Microstructure CNN Classifier |
|
|
| A CNN model for classifying ultra-high carbon steel (UHCS) microstructures from microscopy images. |
|
|
| ## Model Description |
|
|
| Trained on the UHCS Microstructure dataset (Kaggle). Classifies grayscale microscopy images into 4 classes: |
| - spheroidite |
| - network |
| - pearlite |
| - martensite |
|
|
| ## Architecture |
|
|
| - 3 convolutional blocks (16/32/64 filters) |
| - MaxPooling after each block |
| - Fully connected layers (16384 -> 256 -> 4) |
| - Dropout (p=0.5) |
| - Input size: 128x128 grayscale |
|
|
| ## Performance |
|
|
| | Model | Test Accuracy | |
| |---|---| |
| | Logistic Regression (baseline) | 51.3% | |
| | **CNN** | **84.7%** | |
|
|
| ## Usage |
|
|
| Model was trained with PyTorch. To load: |
| ```python |
| import torch |
| model = MicrostructureCNN() |
| model.load_state_dict(torch.load("best_model.pth")) |
| model.eval() |
| ``` |
|
|
| ## Dataset |
|
|
| [UHCS Microstructure dataset on Kaggle](https://www.kaggle.com/datasets/sagarupsc/uhcs-microstructure-01) |
|
|
| ## Full Project |
|
|
| Full code and notebook available on [GitHub](https://github.com/xJadzix/microstructure-classification). |
|
|
| --- |
| license: mit |
| language: |
| - en |
| --- |
