Integer Gaussian convolution with cache memory for real-time processing of the Scale Invariant Feature Transform algorithm

The Gaussian smoothing operator is a 2-D convolution operator which is used to blur images and remove detail and noise in the SIFT (scale invariant feature transform) algorithm. For real-time processing of SIFT, we use our integer Gaussian filtering and cache memory managing schemes using SSE instructions. Single instruction multiple data (SIMD) extensions are currently available in new Pentium processors. We optimize the integer Gaussian filter mask for better precision in key-points detection and compare the result of applying the scheme with those obtained by using the floating processing technique. We apply our scheme to various kinds of images and measure the effectiveness.