Multiscale Image Analysis for the Automated Localisation of Taxonomic Landmark Points and the Identification of Species of Parasitic Wasp

  • Paul Angel

    Student thesis: Doctoral Thesis


    Automating the identification of biological specimens from 2D image data poses difficult problems given the natural variation, specimen damage and background clutter that can exist. The tools used by taxonomists tend to be manual or semi-automated, where the operator locates salient image features from which the system automatically derives taxonomic measurements for identification. Fully automating the extraction of taxonomic features and the subsequent identification task would allow for more robust and accurate identification and provide tools for users in the field who do not possess expert knowledge. This work focuses on the automatic localisation of taxonomic landmark points and the identification of species of parasitic wasps of the order Hymenoptera using SEM images of their heads. These images present significant analysis problems. Image feature extraction techniques investigated to solve this problem
    include deformable contour models, texture analysis and the Mallat wavelet transform. Deformable contour models perform poorly given the textural clutter in the images while texture analysis techniques introduce correlated noise into the segmented image, which can reduce landmark localisation accuracy to 25%. The wavelet transform overcomes this problem by filtering textural clutter at larger scales of analysis. A novel technique is presented which recombines the wavelet transform to create a single contour map where textural clutter is filtered out. This is based on the interaction between edge events which is calculated within a region of interest (ROI) that expands as the scale decreases. In configuring the ROI, a balance must be achieved between filtering textural clutter and eroding salient contours. The landmark localisation accuracy is directly related to this ROI expansion. This represents the main contribution to knowledge. A fast expansion at the high end of the scale range results in a
    landmark localisation accuracy of 95%. Applying these landmarks to a neural network classifier results in a 91% correct identification rate. This represents a significant improvement over the 65% identification rate obtained by taxonomists and is robust to landmark displacement as a result of contour erosion.
    Date of AwardOct 1999
    Original languageEnglish

    Cite this