Binary operator design by k-nearest neighbor learning with application to image resolution increasing

In a typical office environment, heterogeneous devices and software, each working in a different spatial resolution, must interact. As a result resolution conversion problems arise frequently. This paper addresses the spatial resolution increasing of binary images and documents (e.g,, conversion of a 300-dots per inch [dpi] image into 600 dpi). A new, accurate and efficient solution to this problem is proposed. It makes use of the k-nearest neighbor learning to design automatically a windowed zoom operator starting from pairs of in-out sample images. The resulting operator is stored in a look-up table, which is extremely fast computationally and therefore fit for real-time applications. It Is useful to know a prior! the sample complexity (the quantity of training samples needed to get, with probability 1-delta, an operator with accuracy epsilon). We use the probably approximately correct (PAC) learning theory to compute sample complexity, for both noise-free and noisy cases. Because the PAC theory yields an overestimated sample complexity, the statistical estimation is used to estimate, a posteriori, a tight error bound. The statistical estimation is also used to show that the k-nearest neighbor learning has a good inductive bias that allows reduction of the quantity of training sample images needed. (C) 2001 John Wiley & Sons, Inc. Int J Imaging Syst Technol, 11, 331-339, 2000.