A two-step neural-network based algorithm for fast image super-resolution

摘要

We propose a novel, learning-based algorithm for image super-resolution. First, an optimal distance-based weighted interpolation of the image sequence is performed using a new neural architecture, hybrid of a multi-layer perceptron and a probabilistic neural network, trained on synthetic image data. Secondly, a linear filter is applied with coefficients learned to restore residual interpolation artifacts in addition to low-resolution blurring, providing noticeable improvements over lens-detector Wiener restorations. Our method has been evaluated on real visible and IR sequences with widely different contents, providing significantly better results that a two-step method with high computational requirements. Results were similar or better than those of a maximum-a-posteriori estimator, with a reduction in processing time by a factor of almost 300. This paves the way to high-quality, quasi-real time applications of super-resolution techniques.