Feature Extraction and Object Recognition using Conditional Morphological Operators

  • Stephen Rees

    Student thesis: Doctoral Thesis


    This thesis describes the work undertaken on morphological operators for feature extraction from, and segmentation and recognition of, objects within single 2-D images under loosely controlled conditions.

    The novel aspects of the work include the development of a conditional morphological operator, the RJ operator, providing a direct measure of the occupancy of one set by another. This was then applied to the direct extraction of structural features from the intensity map in greyscale images, and to the recognition of objects within images using these features. More complex algorithms for feature identification and object recognition, including a mostly hit, mostly miss transform (MHMMT) and a multiscale structural analysis were developed, using occupancy as the metric. The performance and characteristics of these methods were investigated, using a symmetrical probe as the main tool for analysis and manufactured and natural objects as test pieces.
    Structural features were used as local descriptors of objects. These were extracted by four methods: edge following, chain coding and curvature estimation; direct probing with the R operator and templates; direct probing with the MHMMT; and a generalised R analysis, a multiscale intersection of R operator templated results. The selectivity of the techniques varied, the MHMMT producing the greatest rejection of data. The generalised R analysis produced the most accurate location of features.

    Two methods were adopted to interrelate the extracted features. The first produced a sequence of perimeter features, by estimation of their relative rotations about a calculated feature centroid. The second method interrelated the features as a web skeleton, listing the orientations of each feature relative to the others in the set.

    The multivalued function form of the RJ operator was used to identify the specific object from a model library of poses of various objects. Different combinations of the techniques for extraction and modelling were compared. All objects were recognised, and their orientation determined with errors of between 15 and 25 degrees in the worst cases.
    Date of Award1997
    Original languageEnglish

    Cite this