Peter Rockett

Improved sampling of the Pareto-front in multiobjective genetic optimizations by steady-state evolution: a Pareto converging genetic algorithm

Previous work on multiobjective genetic algorithms has been focused on preventing genetic drift and the issue of convergence has been given little attention. In this paper, we present a simple steady-state strategy, Pareto Converging Genetic Algorithm (PCGA), which naturally samples the solution space and ensures population advancement towards the Pareto-front. PCGA eliminates the need for sharing/niching and thus minimizes heuristically chosen parameters and procedures. A systematic approach based on histograms of rank is introduced for assessing convergence to the Pareto-front, which, by definition, is unknown in most real search problems. We argue that there is always a certain inheritance of genetic material belonging to a population, and there is unlikely to be any significant gain beyond some point; a stopping criterion where terminating the computation is suggested. For further encouraging diversity and competition, a nonmigrating island model may optionally be used; this approach is particularly suited to many difficult (real-world) problems, which have a tendency to get stuck at (unknown) local minima. Results on three benchmark problems are presented and compared with those of earlier approaches. PCGA is found to produce diverse sampling of the Pareto-front without niching and with significantly less computational effort

Boosted key-frame selection and correlated pyramidal motion-feature representation for human action recognition

In this paper we propose a novel method for human action recognition based on boosted key-frame selection and correlated pyramidal motion feature representations. Instead of using an unsupervised method to detect interest points, a Pyramidal Motion Feature (PMF), which combines optical flow with a biologically inspired feature, is extracted from each frame of a video sequence. The AdaBoost learning algorithm is then applied to select the most discriminative frames from a large feature pool. In this way, we obtain the top-ranked boosted frames of each video sequence as the key frames which carry the most representative motion information. Furthermore, we utilise the correlogram which focuses not only on probabilistic distributions within one frame but also on the temporal relationships of the action sequence. In the classification phase, a Support-Vector Machine (SVM) is adopted as the final classifier for human action recognition. To demonstrate generalizability, our method has been systematically tested on a variety of datasets and shown to be more effective and accurate for action recognition compared to the previous work. We obtain overall accuracies of: 95.5%, 93.7%, and 36.5% with our proposed method on the KTH, the multiview IXMAS and the challenging HMDB51 datasets, respectively.

Performance assessment of feature detection algorithms: a methodology and case study on corner detectors

In this paper, we describe a generic methodology for evaluating the labeling performance of feature detectors. We describe a method for generating a test set and apply the methodology to the performance assessment of three well-known corner detectors: the Kitchen-Rosenfeld, Paler et al., and Harris-Stephens corner detectors. The labeling deficiencies of each of these detectors is related to their discrimination ability between corners and various of the features which comprise the class of noncorners.

Robust recognition of scaled shapes using pairwise geometric histograms

The recognition of shapes in images using Pairwise Geometric Histograms has previously been confined to fixed scale shape. Although the geometric representation used in this algorithm is not scale invariant, the stable behaviour of the similarity metric as shapes are scaled enables the method to be extended to the recognition of shapes over a range of scale. In this paper the necessary additions to the existing algorithm are described and the technique is demonstrated on real image data. Hypotheses generated by matching scene shape data to models have previously been resolved using the generalised Hough transform. The robustness of this method can be attributed to its approximation of maximum likelihood statistics. To further improve the robustness of the recognition algorithm and to improve the accuracy to which an objects location, orientation and scale can be determined the generalised Hough transform has been replaced by the probabilistic Hough transform.

An improved rotation-invariant thinning algorithm

Ahmed and Ward [Sept. 1995] have recently presented an elegant, rule-based rotation-invariant thinning algorithm to produce a single-pixel wide skeleton from a binary image. We show examples where this algorithm fails on two-pixel wide lines and propose a modified method which corrects this shortcoming based on graph connectivity.

The training of neural classifiers with condensed datasets

In this paper we apply a k-nearest-neighbor-based data condensing algorithm to the training set of multilayer perceptron neural networks. By removing the overlapping data and retaining only training exemplars adjacent to the decision boundary we are able to significantly speed the network training time while achieving an undegraded misclassification rate compared to a network trained on the unedited training set. We report results on a range of synthetic and real datasets that indicate that a training speed-up of an order of magnitude is typical.

Feature Extraction Using Multi-Objective Genetic Programming

A generic, optimal feature extraction method using multi-objective genetic programming (MOGP) is presented. This methodology has been applied to the well-known edge detection problem in image processing and detailed comparisons made with the Canny edge detector. We show that the superior performance from MOGP in terms of minimizing the misclassification is due to its effective optimal feature extraction. Furthermore, to compare different evolutionary approaches, two popular techniques PCGA and SPGA have been extended to genetic programming as PCGP and SPGP, and applied to five datasets from the UCI database. Both of these evolutionary approaches provide comparable misclassification errors within the present framework but PCGP produces more compact transformations.

A generic optimising feature extraction method using multiobjective genetic programming

In this paper, we present a generic, optimising feature extraction method using multiobjective genetic programming. We re-examine the feature extraction problem and show that effective feature extraction can significantly enhance the performance of pattern recognition systems with simple classifiers. A framework is presented to evolve optimised feature extractors that transform an input pattern space into a decision space in which maximal class separability is obtained. We have applied this method to real world datasets from the UCI Machine Learning and StatLog databases to verify our approach and compare our proposed method with other reported results. We conclude that our algorithm is able to produce classifiers of superior (or equivalent) performance to the conventional classifiers examined, suggesting removal of the need to exhaustively evaluate a large family of conventional classifiers on any new problem.

A generic multi-dimensional feature extraction method using multiobjective genetic programming

In this paper, we present a generic feature extraction method for pattern classification using multiobjective genetic programming. This not only evolves the (near-)optimal set of mappings from a pattern space to a multi-dimensional decision space, but also simultaneously optimizes the dimensionality of that decision space. The presented framework evolves vector-to-vector feature extractors that maximize class separability. We demonstrate the efficacy of our approach by making statistically-founded comparisons with a wide variety of established classifier paradigms over a range of datasets and find that for most of the pairwise comparisons, our evolutionary method delivers statistically smaller misclassification errors. At very worst, our method displays no statistical difference in a few pairwise comparisons with established classifier/dataset combinations; crucially, none of the misclassification results produced by our method is worse than any comparator classifier. Although principally focused on feature extraction, feature selection is also performed as an implicit side effect; we show that both feature extraction and selection are important to the success of our technique. The presented method has the practical consequence of obviating the need to exhaustively evaluate a large family of conventional classifiers when faced with a new pattern recognition problem in order to attain a good classification accuracy.

The Bayesian Operating Point of the Canny Edge Detector

We have investigated the operating point of the Canny edge detector which minimizes the Bayes risk of misclassification. By considering each of the sequential stages which constitute the Canny algorithm, we conclude that the linear filtering stage of Canny, without postprocessing, performs very poorly by any standard in pattern recognition and achieves error rates which are almost indistinguishable from a priori classification. We demonstrate that the edge detection performance of the Canny detector is due almost entirely to the postprocessing stages of nonmaximal suppression and hysteresis thresholding