This is a project of examining various solutions for a Fingerprint Inpainting task (can be found here: https://paperswithcode.com/dataset/fingerprint-inpainting-and-denoising)
A big inspiration was the following article (along with the paper): "Fingerprint Denoising and Inpainting using Fully Convolutional Networks"
| Model | SSIM | PSNR |
|---|---|---|
| Custom Conv + Dense | 0.0997 | 11.7853 |
| Custom Autoencoder | 0.3563 | 13.9658 |
| U-Net (MSE loss) | 0.6989 | 17.0378 |
| U-Net (MS-SSIM loss) | 0.3631 | 12.7265 |
| U-Net (MAE loss) | 0.2409 | 9.0918 |
| Pix2Pix GAN | 0.6266 | 14.3752 |
This model was a sort of "foolish first attempt" at solving the task (as is evident from the architecture)
Architecture:
- 3 x [Conv2D + MaxPooling2D]
- Flatten
- Dense + Dense (IMAGE_HEIGHT x IMAGE_WIDTH) units
- Reshape to (IMAGE_HEIGHT x IMAGE_WIDTH)
Results:
This models implements a fully convolutional autoencoder-like architecture
Architecture:
- Encoder
- 3 x [Conv2D + MaxPooling2D]
- Decoder
- 3 x [Conv2D + UpSampling2D]
- Conv2D
Results:
The U-Net architecture is well known and very popular for image-to-image tasks - for us it turned out to be also the best-performing.
Results:
This model is based on the Pix2pix model: https://www.tensorflow.org/tutorials/generative/pix2pix#build_the_generator
It is a Conditional Generative Adversarial Networks with a U-Net-type model as the generator.
As you can observe on the result below, the discriminator we trained was too weak - therefore the generator gets away with generating real-looking, but completely untrue images (look at the fingerprint lines in detail)
Results:
[TODO]