Page 203 -
P. 203

Section 6.2  Pooled Texture Representations by Discovering Textons  171


                                                                    Rectified           Summarized


                                                  Vertical stripes
                                                                                      Light stripes
                                                                      Positive response  Dark background


                                  Image
                                                                                       Dark stripes
                                                                                       Light background
                                                                      Negative response

                                                Horizontal stripes
                                                                                        Light stripes
                                                                      Positive response  Dark background




                                                                                       Dark stripes
                                                                      Negative response  Light background

                            FIGURE 6.7: Filter-based texture representations look for pattern subelements such as
                            oriented bars. The brick image on the left is filtered with an oriented bar filter (shown as
                            a tiny inset on the top left of the image at full scale) to detect bars, yielding stripe responses
                            (center left; negative is dark, positive is light, mid-gray is zero). These are rectified (here
                            we use half-wave rectification) to yield response maps (center right; dark is zero, light
                            is positive). In turn, these are summarized (here we smoothed over a neighborhood twice
                            the filter width) to yield the texture representation on the right. In this, pixels that have
                            strong vertical bars nearby are light, and others are dark; there is not much difference
                            between the dark and light vertical structure for this image, but there is a real difference
                            between dark and light horizontal structure.


                            the steerable pyramid of Simoncelli and Freeman (1995a). Code for these filters is
                            available at http://www.cns.nyu.edu/ ~ eero/steerpyr/.

                     6.2 POOLED TEXTURE REPRESENTATIONS BY DISCOVERING TEXTONS

                            A texture is a set of textons that repeat in some way. We could find these textons
                            by looking for image patches that are common. An alternative is to find sets of
                            texton subelements—that is, vectors of filter outputs—that are common (if textons
                            are repeated, then so are their subelements). There are two important difficulties
                            in finding image patches or vectors of filter outputs that commonly occur together.
                            First, these representations of the image are continuous. We cannot simply count
                            how many times a particular pattern occurs, because each vector is slightly different.
                            Second, the representation is high dimensional in either case. A patch around a
   198   199   200   201   202   203   204   205   206   207   208