Autoencoder-Huffman Ensemble Model for Image Compression

Autoencoders are deep learning tools that excel at compression and noise reduction. They work by first passing input data through an encoder to obtain a Latent Space Representation (LSR), which captures essential information. This LSR is then used by the decoder, the other half of the autoencoder network, to reconstruct the data. Because the LSR has a low number of neurons, it retains only the most critical information, effectively eliminating much of the noise and redundancy from the input. This paper investigates the use of an Autoencoder-Huffman ensemble for compression. By incorporating a Huffman module alongside the original autoencoder network, the compression is achieved with minimal loss of information. The method involves two levels of compression: one from the autoencoder to the LSR and another from the LSR to Huffman encoding. To assess its performance, the proposed ensemble model is tested on three different datasets: MNIST, CIFAR-10, and Fashion MNIST. The results show that the ensemble model achieves high Average Compression Ratios (ACRs) for all three datasets, both with and without Huffman code overhead. For MNIST, CIFAR-10, and Fashion MNIST, the ACRs without overhead are 15.654, 24.230, and 15.171, respectively. With overhead, the ACRs are 5.510, 8.686, and 5.313, respectively, using LSRs of 100, 90, 80, and 70.