gperdrizet/denoising_autoencoder

Image denoising autoencoder

A convolutional autoencoder trained to remove Gaussian noise from 256×256 RGB images. The model learns the manifold of clean images and projects noisy inputs back onto this learned space.

Model description

This model removes Gaussian noise from corrupted images by learning to map noisy images back to their clean versions. During training, the autoencoder learns the distribution of clean images, enabling it to filter out noise while preserving image content.

Architecture:

• Encoder: Convolutional layers with downsampling (256×256×3 → 1024)
• Decoder: Transposed convolutional layers with upsampling (1024 → 256×256×3)
• Activation: LeakyReLU and Sigmoid
• Normalization: Batch normalization

Training noise level:

• σ=25 (on 0-255 scale) Gaussian noise
• Approximately 0.098 on [0,1] normalized scale

Performance:

• Target PSNR improvement: Noisy <25 dB → Denoised >28 dB
• Target SSIM: >0.85

Intended use

This model is designed for educational purposes to demonstrate how autoencoders can learn to denoise images by understanding the structure of clean data.

Use cases:

• Understanding autoencoder denoising
• Learning about noise removal techniques
• Exploring how neural networks learn image manifolds
• Teaching AI/ML concepts in bootcamps

Training data

Trained on DF2K_OST, a combined dataset of high-quality images from:

• DIV2K (800 images)
• Flickr2K (2,650 images)
• OutdoorSceneTraining (10,424 images)

All images resized to 256×256 pixels using Lanczos resampling. Gaussian noise (σ=25) added during training.

Training details

Hyperparameters:

• Optimizer: Adam (lr=1e-3)
• Loss function: Mean Squared Error (MSE)
• Batch size: 8
• Epochs: Up to 100 (with early stopping)
• Train/validation split: 90/10
• Noise level: σ=25

Callbacks:

• Early stopping (patience=5, monitoring validation loss)
• Learning rate reduction (factor=0.5, patience=3)
• Model checkpoint (best validation loss)

Hardware:

• Single NVIDIA GPU with memory growth enabled

How to use

import numpy as np
from tensorflow import keras
from huggingface_hub import hf_hub_download

# Download model
downloaded_model = hf_hub_download(
    repo_id='gperdrizet/denoising_autoencoder',
    filename='models/denoising_ae.keras',
    repo_type='model'
)

# Load model
autoencoder = keras.models.load_model(downloaded_model)

# Denoise images
denoised = autoencoder.predict(noisy_images)  # images shape: (N, 256, 256, 3)

For complete examples, see the training notebook.

Limitations

• Fixed input size (256×256 RGB images)
• Trained on σ=25 noise level (may not generalize well to other noise levels)
• Not optimized for non-Gaussian noise
• Slower than traditional denoising filters
• Educational model, not production-ready

Generalization

The model was trained on σ=25 Gaussian noise but can handle different noise levels with varying effectiveness:

• Lower noise (σ<25): Excellent performance
• Similar noise (σ≈25): Best performance
• Higher noise (σ>25): Reduced effectiveness

Project repository

Full code, training notebooks, and interactive demo: gperdrizet/autoencoders

Citation

If you use this model for educational purposes, please reference the project repository.