This project done by 3 of us demonstrates the use of diffusion-based generative modeling to learn structural patterns from biomaterial microstructure images. The model is trained on grayscale structural data and generates new microstructure-like patterns. The generated outputs are evaluated and interpreted using similarity metrics and structural analysis techniques.
- Microstructure images are preprocessed into grayscale format
- Images are resized to 64×64 resolution
- Pixel values are normalized for training
- A simplified UNet-based convolutional model is used
- Encoder-decoder structure learns structural features
- Model is trained to remove noise from input images
- Gaussian noise is added to images
- Model learns to reconstruct original structures
- Optimization is performed using Mean Squared Error loss
Edge detection highlights structural boundaries and repeating patterns, indicating that the model captures spatial features present in the dataset.
The frequency spectrum shows dominant low-frequency components, suggesting the presence of distributed structural textures.
The histogram indicates variation in structural density and grayscale intensity across the generated image.
The distribution of orientations shows structural features spread across multiple directions, indicating isotropic characteristics.
The orientation map visualizes spatial variation of structural directions within the generated microstructure.
| Metric | Value |
|---|---|
| Average SSIM | 0.0092 |
| MSE | 0.321 |
| PSNR | 4.93 dB |
Interpretation
The low similarity values indicate that the generated images are not replicas of specific training samples but represent new structures derived from the learned distribution.
- The model captures global structural patterns rather than exact replicas
- Generated images exhibit microstructure-like textures
- Structural analysis confirms pattern learning
- Orientation and frequency analyses validate spatial characteristics
- Python
- PyTorch
- NumPy
- OpenCV
- Matplotlib
YOUVASHREE K
RAMKUMAR R
PRAGALYA M
This project demonstrates that diffusion-based models can learn and generate structural patterns from biomaterial datasets. The combination of generative modeling and structural interpretation provides a meaningful approach for analyzing material structures using machine learning.