General Structure of an Machine Learning Method for Compression of Images

The article proposes an architecture of based on machine learning (ML) method for compressing digital images and two-dimensional signals. The method uses a quadtree-based representation of the two-dimensional signal. This representation allows you to use the characteristics of the different quadtree floors to tuning the ML-methods. The representation also allows you to predict two-dimensional signal samples based on less detailed quadtree floors and encode prediction deviations. We use ML-methods at all stages of the compression method. In particular, we use ML-based super-resolution and embedding algorithms for digital two-dimensional signals to predict signal samples: convolutional and generative adversarial neural networks, as well as autoencoders. We perform computational experiments to investigate the efficiency of ML-based prediction at the most detailed floor of the quadtree-based representation of the two-dimensional signal during compression. We experimentally prove that the studied ML-based predictor has a significant advantage in natural digital two-dimensional signals.