- This is a course project titled 'Image Compression and Reconstruction Using VAE' done under the guidance of Prof. Asim Tewari with a six-person team in the Statistical Machine Learning and Data Mining Course(ME-781), IIT Bombay.
- This project was focused on data compression using VAE(Variational Autoencoders) to learn compact representations of data while retaining essential information.
- Reduce the size of input data while preserving its essential features.
- Ensure that the compressed data retains a high level of quality.
- A VAE model using TensorFlow and Keras on the MNIST dataset was implemented.
- VAE is a type of generative model that can learn a probabilistic mapping between the input data and a latent space.
- Encoder of VAE was used to compress images and Decoder to decompress images.
- Trained on open sourced MNIST handwritten digits dataset (available in keras.datasets)
- Data Preprocessing: Normalized pixel values to the range [0, 1] then flattened the 28x28 images into 1D arrays.
- Model Compilation: Compiled the VAE model using the Adam optimizer.
- Training Loop: Fitted the VAE model to the training data using a batch size of 64 and trained for 20 epochs. Trained on 6000 images and tested on 1000 images
- Compression and Decompression: Used the trained encoder to compress the test data into the latent space.
- Reconstruction: Used trained decoder to reconstruct the data from the compressed latent space.
- Visualisation: Visualised the reconstructed images for qualitative assessment.
- Images before passing to the model:
- Reconstructed images by decoder:
