Multi-Resolution Recursive Peak Tracking: An Algorithm for Shape Representation and Matching

This report describes ,a rcprcsentation for gray level patterns in digitized images based on intensity peaks at multiple resolutions. The Multi-resolution Recursive Peak Tracking (MRRfrF) algorithm a method for locating a representative subset of the multi-resolution band-pass peaks for a form in an image. The term form is meant o denote a region in an image which is distinct from the background. The representation is shown to be tolerant to object translation and scaling. The MRRPT algorithm is based on a fixed-pattern peak search using size-scaled square bandpass filters. The search is limited to the surface of a rectangle which has sides parallel to the image axes and encloses a form in IRe image, thus reducing the computations required. The algorithm is best suited to the representation of isolated forms. In order to evaluate the descriptive content of the information extracted, the effectiveness of the algorithm for classification of objects using training-based models i examined. Ahi-level similarity measure is used allowing objects to be classified into groups of identical as well as similar gray scale shape. The algorithm was able to successfully classify 92% of 24 gray-level images of tools into 6 groups of exact matches, and 90% of 20 puzzle pieces into 5 groups of similar matches, The remaining 8% of the tools were correctly identified, but were classified as s#nilar matches. The classification performance of the MRRPT algorithm is compared to other methods of binary and gray-level image shape description. The cost of the algorit2am is dependent upon the content of the image and the size of the form in the image. The comple~:ity is shown to be approximately linear in the number of pixels for a fixed number of levels of resolution. The cost and complexity of the algorithm are compared to other methods of image shape :~epresentation. Chapte r