Finding complex patterns using template matching

Abstract

Complex patterns that combine pixels of different brightness and colour in a fixed spatial layout pose a challenge for the objectoriented approach to image analysis. This is because the pixels that make up the pattern cannot readily be linked together into an object as they do not share spectral or textural properties. The goal of this study was to determine whether template matching techniques can be used to identify those patterns without a prior segmentation of the image. We studied three different scenarios that pose specific challenges: the application of template matching is most straightforward in (1) the detection of symbols in scanned maps where different occurrences of a given pattern strongly resemble one another; most notably, symbols always have the same size and orientation. (2) The detection of trees in aerial images of oil palms is more challenging, since the characteristic tree pattern exhibits significant variations from tree to tree. Finally, in the (3) detection of pavement markings in aerial images objects could appear at arbitrary orientations in the image. Our results show that template matching successfully identifies many target regions. Once target locations are known, objects can be created and the full power of object-based image analysis approaches can be applied to eliminate remaining false positives.