NeuralNetwork-Based-Generic-Compression

Generic compression using autoencoder

---

Installation

Python 3.X.X is required for this to work.

1. Create your python(v3) virtual environment
2. Install the necessary packages

pip install tensorflow==2.1.0
pip install keras==2.3.1
pip install opencv-python==4.2.0.32
pip install Pillow==7.0.0
pip install image==1.5.28
pip install noisereduce
pip install numpy
pip install matplotlib

3. Download the following files from this github repo:

• main.py
• EncoderDecoderAudio.py
• EncoderDecoderImage.py
• audio_autoencoder.model
• image_autoencoder.h5

OR

• Download this .zip file --> Click Here

4. Put the downloaded files in a single folder

---

Usage

Open your CMD and navigate to the installation folder from step 4 (from the Installation section)

Note: the model detects the filetype (image/audio) automatically, you don't have to specify.

---

Encode Mode

To Encode (compress), use the following command:

python main.py encode [input_file_path] [compressed_file_path]

Examples:
python main.py encode myimage.png mycompressed
python main.py encode myaudio.wav mycompressed

Note: You are required to include the input file extension but not the compressed file.

---

Decode Mode

To Decode (decompress), use the following command:

python main.py decode [compressed_file_path] [output_file_path]

Example:
python main.py decode mycompressed my_image_output.png
python main.py decode mycompressed my_audio_output

Note: You are required to include the output file extension for the image output only.

---

Supported formats

Audio:

• .wav

Image:

• .JPEG
• .JPG
• .PNG
• .TIFF

---

Datasets Used

Image Datasets

Note: Not all the content of the datasets were used due to resources limitations. https://www.kaggle.com/evgeniumakov/images4k
http://www.cs.toronto.edu/~kriz/cifar.html
https://www.kaggle.com/hsankesara/flickr-image-dataset
https://www.kaggle.com/vishalsubbiah/pokemon-images-and-types

• All images are processed first using data_generator.py before being used for training.
• All images are cut into 32x32 blocks to match the model's input size.
• Around ~1,000,000 32x32x3 images are used for training. (dataset contains 15,000,000+)

Audio Dataset

• Beatport EDM Key Dataset https://zenodo.org/record/1101082#.XqyLuqgzZPZ

Note: A portion of 2.49 GB Wav files (125 song) of the dataset is used in training.