Quaternionic multilayer bottleneck autoencoder for color image compression

This paper presents a study on the multilayer bottleneck autoencoder applied for thumbnail color image compression task using quaternionic neural nets (QNN). Existing state-of-the-art real-valued autoencoder architectures are weak for reconstruction of images with high-frequency components. However, experiments shows that hypercomplex representation of color and activation functions, gives better results in terms of pixellevel quality metrics PSNR and SSIM. Designed hypercomplex autoencoders are simplest type of the QNN - fully connected quaternionic layers. Investigated one- and two-layer QNN structures for encoder/decoders in compare with real-valued ones provides better PSNR by approx. 2.5 dB and 0.11 SSIM (Real valued: PSNR - 26.36 dB, SSIM - 0.856; Quaternionic: PSNR 31.28 dB, SSIM - 0.951) for 3:1 compression ratio and equal or less number of hyperparameters.