Abstract
Robustness to geometrical transformations such as rotation, scaling and translation (RST) as well as noise are major concerns in computer vision and image analysis. This thesis proposes an effective feature extraction approach using Trace transform and ensembles of classifiers for invariant image identification. The key question in Trace transform is to select the best combinations of the Trace functionals to produce and apply the optimal Triple features with minimum computational cost, which is a challenging task because robustness and computational speed conflict with each other. This challenge poses a series of experimental and analytical discussions outlined into two phases. In the first phase, we propose Evolutionary Trace Transform (ETT) that adopts evolutionary algorithms and Pareto optimality principles to select the best functionals used in Trace transform. To tackle noise, we deliberately inject noisy samples in the example images in the evolutionary training of Trace transform, which is termed Evolutionary Trace Transform with Noise (ETTN). Single-objective and multi-objective optimisation were developed and compared. A one-shot approach is considered, which uses a very small number of examples in the evolutionary training process and applies the extracted features to test the entire dataset. The second phase deals with building classifiers. To complete the identification task, a variant of classifiers were constructed and compared. To further enhance the performance and to increase the reliability of the system, we propose ensembles of classifiers that use multiple Pareto optimal image features. The proposed ensembles take advantage of the diversity inherent in the Pareto optimal features extracted using the ETTN algorithm and empirical results show that on average, ensembles using Pareto optimal features perform much better than traditional classifier ensembles using the same features and data randomisation. Diversity analysis using a number of measures is also considered, indicating that the proposed ensembles consistently produce a higher degree of diversity than traditional ensembles. Furthermore, a tuning process of the Trace transform parameters is conducted to obtain a trade-off between complexity and robustness using two evolutionary multi-objective approaches. In the first approach, two-objective evolutionary algorithms are adopted using the within-class variance and between-class variance as objectives. The second approach adopts three-objective evolutionary algorithms, which consider the computational complexity as a third objective. Two different coding schemes are considered for each approach, which are integer-coding and real-coding schemes. The proposed schemes were compared by conducting experiments on sample images and results showed that the integer-coding scheme presents a better performance compared to the real-coding scheme. Moreover, while the three-objective approach enforces a balance between robustness and computational complexity, without enforcing a minimum acceptable accuracy, features extracted tend to have a lower computational complexity at the expense of the accuracy, compared with the two-objective approach.