Lossless Image Compression Based on Image Decomposition and Progressive Prediction Using Convolutional Neural Networks

This paper presents a lossless image compression method based on the image decomposition and progressive prediction of decomposed images and their coding contexts using convolutional neural networks (CNNs). We first decompose a given input into sub-images by sub-sampling into horizontal and vertical directions. The first sub-image is encoded by an existing non-learning lossless compressor, and the others are encoded progressively using the already encoded ones. While state-of-the-art learning-based encoders proceed with the auto-regressive prediction (pixel-by-pixel raster-scan order prediction) so that the runtime is not practical, our method predicts each sub-image at once and thus executes in practical time. We also design the CNNs to predict pixel values and coding contexts jointly and send the prediction error to the arithmetic encoder for the given coding context. In the case of color input, it is converted to YUV using a reversible transform, and each channel is partitioned in the same way above. Then, all the sub-images are processed progressively from Y to V. Experiments on high-resolution datasets show that the proposed method outperforms all the non-learning codecs and practical-time CNN-based encoders. Compared to learning-based auto-regressive approaches, our method shows comparable results while requiring much less runtime.