Deep Signal Compression with Feature Representation Learning

DNN-based lossy signal compression has achieved substantial progress in recent years. Signal source coding can benefit from learned methods in two ways. Firstly, the data-intense nature of deep signal compression allows a good capture of the probabilistic distribution of feature representations, which leads to efficient entropy coding. Secondly, neural network architectures can provide solutions to feature extraction and representation learning, therefore enabling the elimination of spatial and temporal redundancies by mapping the raw signal to compact feature domains.

This dissertation presents four related works addressing the compression problem of different data formats. The first work introduces a unified method that uses generative adversarial networks to compress speech and images. The compressed signal is represented by a latent vector minimizing a target objective function. The second work presents a deep video coding framework that predicts and compresses video sequences in the latent vector space. The third work addresses the motion pattern adaptability issue that widely exists in video codecs by a blockwise mode ensemble framework. The last work proposes a LiDAR data compression pipeline, which primarily relies on a prediction-based approach to exploit spatial and temporal correlations in range images while providing an octree-based path as a fallback to preserve quality.

CHAIR:  Professor Hun-Seok Kim