Pyramidal Image Compression Based on Machine Learning
Mikhail V. Gashnikov · 2022
We develop a machine learning-based image compression method. We propose the general structure of the method based on detection of uninformative image areas and subsequent pyramidal compression of highly informative areas. The use of image inpainting algorithms during decompression allows us not to compress uninformative areas and not to put the areas in the archive. We use adversarial neural networks for the image inpainting. The irredundant pyramidal representation of the image allows us to compress the layers of this pyramid sequentially. Therefore, we use the structural features of each layer of the pyramid for machine learning. Super-resolution neural network algorithms allow us to improve the efficiency of interpolation of all image pyramid layers. Approximation of image spectral bands allows us to use interband correlations and generalize the proposed compression method to hyperspectral images. We analyze machine learning algorithms and select a specific algorithm for each stage of the proposed compression method. We perform computational experiments to study the effectiveness of individual stages of the proposed compression method in real images. We experimentally prove that the selected machine learning algorithms are effective within the image compression method.