אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

Non uniform filtering is important for many image processing algorithms. However, for large kernel sizes it can become computationally expensive. In this talk we describe two efficient filtering techniques which are independent on kernel size. First we present Cosine Integral Images (CII) which represent a large set of spatial and range filters, based on their frequency decomposition. We apply CII for fast computation of the spatial Gaussian and Gabor kernels, whose complexity is a constant O(1) operations per image pixel. We also improve previous constant time approximations of bilateral filtering.

I n addition we present a simple and very efficient method for Gaussian convolution using running sums along the image rows and columns. In order to approximate the Gaussian kernel we investigate relation between the error function used for kernel approximation and the resulting L2 error on the output image. Based on natural image statistics we propose a quadratic form error measure, and use it to approximate the Gaussian kernel by linear combination of constant functions. This results in very efficient Gaussian filtering method. Our experiments show that the proposed technique is faster than state of the art methods while preserving a similar accuracy.

Joint work with Prof. Michael Werman