There is a serious and increasing taxonomic impediment caused primarily by a lack of trained taxonomists. Computer aided taxonomy (CAT) is one possible way of solving the problem by speeding up and partially automating the identification of taxa. The work described in this thesis develops a new CAT technique which performs semi-automated identification of two groups of insects: hoverflies and bumblebees using images of their wings. A series of image processing algorithms have been designed to extract high quality venation diagrams via a user-friendly software package. Taxon identification of is achieved in two ways – (1) a novel analysis technique based on the venation and relationships between veins using a tree like diagram and tree comparison, and (2) extraction of characteristic features such as cell composition and vein fitting coefficients used as the input parameters of artificial neural networks. Tree-based identification produces 100% recognition accuracy for taxa to tribe level for 9 hoverfly species and 3 bumblebee species (including 2 sub-species). The accurate species-level identification using 2 different types of artificial neural networks (multi-layer perceptron and learning vector quantisation) achieve a recognition rate of 90.9% for hoverflies and 86.7% for bumblebees. These identifications are achieved by learning vector quantization, and the average results for multi-layer perceptrons are much lower at 60% for hoverflies and 30% for bumblebees. Test results prove that this semi-automated technique using structural image processing can provide a friendly used tool for identification between morphologically similar species of insects and can be applied to any groups with transparent wings (e.g. Diptera, Odonata, and other Hymenoptera).
The figure below shows a scanned Diptera wing (Eristalis tenax) and it’s automatically extracted venation.
The novel aspect of this work is on the development of a tree description of the vein relationships. A typical tree for E. tenax is shown below.