Lossless Image Compression with Learned Local and Global Features

Estimating the probability distribution of an image is the key issue in lossless image compression. Though image compression can benefit from both global and local information, few works have been proposed to utilize both in lossless image compression. In this work, we propose to use a neural network for multiscale feature learning, the learned features are used to estimate the distribution of the image in a chain rule. In a further step, we utilize a context model to learn local features from the image. Finally, we combine the multiscale features with local features for image distribution learning. Our work surpasses state-of-the-art learning algorithms and several traditional codecs in several challenging datasets.