This thesis considers two problems in classification -- a field within artificial intelligence. One is the general problem of classifier learning, for which a meta-classification technique called 'virtual attribute subsetting' is developed and tested. The other is object detection, with emphasis on marine creature detection, using the 'Haar Classifier Cascade' method (Viola & Jones, 2001b). Haar Classifier Cascades are built from a feature set of simple rectangular patterns of relative light and dark. Adaptive boosting selects those features that best tell the difference between objects and non-objects. In this thesis, a new cascade confidence measure is proposed, equivalent to the boosting 'margin'; it uses information about how well the cascade features match the image region being classified. Tests on the common application of face detection show that this confidence measure improves detection accuracy. Virtual attribute subsetting is also used to modify the cascade; it further improves accuracy at the expense of classification time. In addition, Haar Classifier Cascades are trained to detect two types of marine animal (fish and seahorses). This requires object detection across a wide range of orientations, so approaches using both image and cascade rotation are compared. Results show that image rotation is more accurate than cascade rotation, and that cascades trained to detect objects over a range of angles should have their training images randomly perturbed over a similar (but not always equal) range. Confidence-based detections are also made and show themselves to be more accurate than binary detection. The confidence-based results sum the confidences from similar detections and show that confidence measurements from multiple Haar Classifier Cascades may be combined effectively. Seahorse detection poses an additional problem: seahorses are too flexible to be found by single cascades in any orientation. To solve this, separate seahorse head and body detectors are trained and their detections matched to create whole seahorse detections. Both designed and learnt matching cost formulae are created and two matching algorithms are implemented to link together head and body detections given a cost measurement. The best of the resulting whole seahorse detectors is more accurate than either of the component part detectors. The confidence measurement and virtual attribute subsetting algorithms make no use of domain knowledge, so should improve the accuracy of most other Haar Classifier Cascades. They also have the unusual property of being applicable to already learnt classifiers.