| --- | |
| title: DeepClean CNN Autoencoder For Image Denoising | |
| emoji: 🎨 | |
| colorFrom: purple | |
| colorTo: blue | |
| sdk: docker | |
| pinned: false | |
| app_port: 7860 | |
| short_description: CNN Autoencoder image denoiser trained on MNIST | |
| # DeepClean CNN Autoencoder for Image Denoising | |
| A deep learning web app that removes noise from handwritten digit images using a **Convolutional Autoencoder** trained on the MNIST dataset. | |
| ## How It Works | |
| Upload a noisy grayscale image (any size it gets resized to 28x28 automatically), and the model reconstructs a clean version. | |
| ### Model Architecture | |
| - **Encoder**: Conv2D(32) -> MaxPool -> Conv2D(16) -> MaxPool -> latent space (7x7x16) | |
| - **Decoder**: Conv2D(16) -> UpSample -> Conv2D(32) -> UpSample -> Conv2D(1, sigmoid) | |
| ### Performance | |
| | Metric | Value | | |
| |--------|-------| | |
| | Test Accuracy | 87.56% | | |
| | F1 Score | 0.8923 | | |
| | Test Loss | 0.1234 | | |
| ### Dataset | |
| - **MNIST** Handwritten Digits | |
| - 60,000 training samples / 10,000 test samples | |
| - Gaussian noise (factor = 0.5) added during training | |
| ## Tech Stack | |
| - TensorFlow / Keras | |
| - Flask | |
| - Pillow | |
| - Docker (Hugging Face Spaces) | |