|dc.description.abstract||In New Zealand all land mammals (except bats) were introduced through human migration. These animals mostly came to the country as pets, or were introduced for economic reasons. Currently there is a huge problem with these introduced species. Currently species recognition is one of the key methods to understand and control certain animal in an area. This thesis investigates three existing techniques for recognising possum and cat. These techniques Eigenface, Fisherface and Support Vector Machine (SVM) are a novel application in animal recognition domain. When these techniques are trialled with possum and cat images, recognition rates are not acceptable for application. To improve the recognition rate a few methods are investigated. They are different colour schemes, different image resolutions and finally an error weight-based algorithm to measure the distance between average image and test image for the Eigenface technique. This developed technique produced favourable results for possums and cats detection. This new technique is compared with typical Eigenface, Fisherface and SVM techniques to investigate its performance. To further investigate, the developed algorithm is tested with dogs to check the performance with other animal species. Results show that there is an acceptable separation, with this multiclass problem. The performance (class separation) is analysed using the Receiver Operating Characteristic (ROC) method. This method provides a unique way to compare the class separation of the above three techniques.
Finally, a ROC-based feature selection method is developed to use with Principal Component Analysis (PCA), Fisherface and SVM techniques. This new technique helps to find the optimal dataset for training the above techniques. Lastly one of the main challenges to this investigation was the limited availability of training images for the algorithms. Due to the prohibitive cost of animal ethics approval, the training images were obtained from the internet. Hence all developed identification algorithms were optimised for small training sets.||en_NZ