Rotation, scale and translation-invariant segmentation-free grayscale shape recognition using mathematical morphology
Sidnei Alves de Araújo, Hae Yong Kim · 2007
Rotation, scale and translation-invariant grayscale shape recognition is an important problem in computer vision. However, to our knowledge, all practical techniques in the literature first “simplifies” the image using operations like segmentation/binarization and detection of edges and corner points. Some approaches that achieve RSTinvariance by detecting interest points and edges include: curvature scale space [1], orientation code histograms [2], geometric hashing [3], and generalized Hough Transform [4]. Other techniques, like [5], first binarizes the image, isolates the shapes, normalizes their area and extracts some RST-invariant features. The most commonly used RST-invariant features are Hu’s moments [6] and Zernike’s moments [7]. However, these “image simplifying operations” throw away the rich grayscale information, are noisesensitive and prone to errors. Thus, a segmentationfree approach is attractive. Segmentation-free approaches were proposed to recognize printed character [8] and handwritten numeral string [9], but they are not RST-invariant. Mathematical morphology has been used successfully in many works related to shape recognition, for example [10]. In this paper, we present a RST-invariant, segmentation-free graylevel shape recognition method using mathematical morphology approach. It is composed of three steps. In the first and second steps, filters based on dilations and erosions prune out the pixels that have no chance of matching the query shape. The third step makes use of the conventional template matching to recognize the query shape.