Face recognition is a key human brain function as faces convey a wealth of information about a person's mood, intentions, interest, health, direction of gaze, intelligence and trustworthiness, among many factors. Previous studies gained from behavioural, functional magnetic resonance imaging (fMRI), electroencephalographic (EEG) and MEG studies have shown that face processing involves activity in many specialised areas of the brain, which are known collectively as the face processing system. The aim of this thesis has been to develop, apply and assess a novel technique of analysis in order to gain information about the face processing system. The new technique involves using Independent Components Analysis (ICA) to identify significant features in the data for each subject and then using k-means clustering to aggregate results across subjects. A key feature of this new technique is that it does not impose a priori assumptions on the localisation of the face processing system in either time or space, and in particular does not assume that the latency of evoked responses is the same between subjects. The new technique is evaluated for robustness, stability and validity by comparing it quantitatively to the well established technique of weighted Minimum Norm Estimation (wMNE). This thesis describes a visually evoked response experiment involving 23 healthy adult subjects in which MEG data was recorded as subjects viewed a sequence of images from three categories: human faces, monkey faces or motorbikes. This MEG data was co-registered with a standard head model (the MNI30S brain) so that inter-subject comparisons could be made. We identify six clusters of brain activity with peak responses in the latency range from lOOms to 3S0ms and give the relative weighting for each cluster for each the three image categories. We use a bootstrap technique to assess the significance of these weightings and find that the only cluster where the human face response was significantly stronger than the motorbike image response was a cluster with peak latency of l72ms, which confirms earlier studies. For this cluster the response to monkey face images was not significantly different to the human face image response at the 99% confidence level. Other significant differences between brain response to the image categories are reported. For each cluster of brain activity we estimate the activity within each labelled region of the MNI30S brain and again use a bootstrap technique to determine brain areas where activity is significantly above the median level of activity. In a similar way we investigate whether activity shows hemispherical bias by reporting the probability that we reject the null hypothesis that the left and right hemispheres have the same level of activation. For the clu~ter with peak latency at 172ms mentioned above we find that the response is right lateralised, which again confirms earlier studies. In addition to this information about the location of brain activity, the techniques used give detailed information about time evolution (and sequencing) that other techniques such as fMRI are unable to provide. This time evolution of the clusters shows some evidence for priming activity that may give advance notice of the importance of a new visual stimulus, and also some support for a theory of anterior temporal lobe involvement in face identification (Kriegeskorte2007). We also describe activity that could be attributed to executive systems and memory access,'

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.

Automatically extracting interesting objects from videos is a very challenging task and is applicable to many research areas such robotics, medical imaging, content based indexing and visual surveillance. Automated visual surveillance is a major research area in computational vision and a commonly applied technique in an attempt to extract objects of interest is that of motion segmentation. Motion segmentation relies on the temporal changes that occur in video sequences to detect objects, but as a technique it presents many challenges that researchers have yet to surmount. Changes in real-time video sequences not only include interesting objects, environmental conditions such as wind, cloud cover, rain and snow may be present, in addition to rapid lighting changes, poor footage quality, moving shadows and reflections. The list provides only a sample of the challenges present. This thesis explores the use of motion segmentation as part of a computational vision system and provides solutions for a practical, generic approach with robust performance, using current neuro-biological, physiological and psychological research in primate vision as inspiration.

This thesis develops a registration algorithm specifically for diffusion-tensor (DT) images. The proposed approach matches the tensor orientations to find the registration transformation. Early results show that local optimisation does not find the global minimum in registration of DT-MR brain images. Therefore, a global optimisation registration technique is also implemented. This thesis proposes several new similarity measures for DT registration and provides a comparison of them along with several others previously proposed in the literature. The thesis also proposes several new performance evaluation measures to assess registration quality and develops a performance evaluation framework that uses directional coherence and landmark separation. Experiments with direct optimisation demonstrate increased local minima in tensor registration objective functions over scalar registration. Using registration with global optimisation, this thesis compares the performance of scalar-derived similarity measures with those derived from the full tensor. Results suggest that similarity measures derived from the full tensor matrix do not find a more accurate registration than those based on the derived scalar indices. Affine and higher-order polynomial registration is not reliable enough to make a firm conclusion about whether diffusion tensor orientation matching improves the accuracy of registration over registration algorithms that ignore orientation. The main problem preventing a firm conclusion is that the local minima problem persists despite the use of global optimisation, causing poor registration of the regions of interest.