Harris and Laplacian Region Detectors

This section discusses the procedure for extracting scale-adapted local regions, i.e., interest points equipped with characteristic scales. The Laplacian detector extracts image regions whose locations and characteristic scales are given by scale-space maxima of the Laplace operator. The Harris detector uses the same operator for scale selection, but finds the locations of interest points as the local maxima of a “cornerness” measure based on the second moment matrix. Let us begin by summarizing Lindeberg’s procedure for spatial and scale selection using the Laplace operator [2]. Let I(x) denote the image intensity as a function of position. To simplify the presentation, we will treat I(x) as a continuous differentiable function. We can form a basic linear scale space by considering all possible images that result from convolving I(x) with an isotropic Gaussian of standard deviation σ: