Orthogonalization and Parameterization of Convolutional Kernels in Machine Learning for Image and Video Compression

Abstract

We study orthogonalization and parametrization of convolutional filters within the framework of the image and video compression method based on machine learning. We use the convolutional filters to interpolate less sparse video frame meshes based on sparser video frame meshes. We consider superresolution neural networks and decision trees as machine learning algorithms at the interpolation stage. Decision trees adaptively select an interpolating function from a predefined set of convolutional filters with parameterized orthogonal weights. The use of adaptive functions can significantly improve the accuracy of interpolation. Optimization of machine learning algorithms makes it possible to use the adaptability of interpolators in the most efficient way. We use orthogonalization and parametrization of convolution filter weights to increase the efficiency of the machine learning interpolation algorithm, which in turn leads to an increase in the efficiency of the image and video compression method in general. Computational experiments demonstrate the advantage of the proposed algorithm in real videos.