Autoencoder - Image Compression

Context

Developed an autoencoder neural network for image compression on the Fashion-MNIST dataset. The project focused on optimizing the architecture to maximize Structural Similarity Index (SSIM) while minimizing Mean Squared Error (MSE).

Technologies Used

• Framework: PyTorch
• Dataset: Fashion-MNIST (60,000 training images)
• Libraries: NumPy, Matplotlib, Scikit-learn
• Metrics: SSIM, MSE, compression ratio

Implementation

Architecture:

• Encoder: Convolutional layers for feature extraction
• Latent space: Compact bottleneck representation
• Decoder: Transposed convolutions for image reconstruction

Training:

• Loss function: Combination of MSE and SSIM
• Optimizer: Adam
• Regularization: Dropout and BatchNorm

The autoencoder learns to compress 28×28 grayscale images into a compact latent representation and reconstruct them with minimal information loss.

Results

• Achieved excellent compression ratio
• High SSIM score indicating good perceptual quality
• Low MSE demonstrating accurate reconstruction
• Visual quality: Near-perfect reconstruction for most samples

Challenges & Learnings

• Finding the optimal latent space dimension (trade-off between compression and quality)
• Tuning loss functions to prioritize perceptual similarity
• Understanding the learned features in the latent space