Generative Models
Fashion Autoencoder
Analysis of Autoencoder use cases for clothing images: dimensionality reduction, denoising, and generation.
Course: Neural NetworksCo-authors: Diego Quezada
Objectives
- 1Compare fully connected versus convolutional autoencoder architectures for image compression.
- 2Evaluate autoencoders for three use cases: dimensionality reduction, denoising, and image generation.
- 3Analyze how latent space dimensionality affects reconstruction quality.
Conclusions
- Convolutional autoencoders achieve 40% lower reconstruction error than fully connected architectures.
- A 32-dimensional latent space preserves essential visual features while achieving 24x compression.
- Denoising autoencoders effectively remove Gaussian noise but tend to over-smooth fine details.
- Image generation from random latent samples produces recognizable silhouettes but lacks texture detail.
Technologies
- TensorFlow
- Keras
- NumPy
- Matplotlib