Constraint directed learning for unsupervised image sequence segmentation

In applications such as the segmentation of infrared images, a classifier can be used to map features of a pixel's neighbourhood to a discrete class. By applying the classifier at every position a segmentation is obtained. An unsupervised classifier can learn by clustering the input vectors in feature space. Clusters can then be regarded as classes. However such schemes do not automatically make use of the spatial context associated with feature vectors. Spatial context can aid the formation of clusters. For example, pixels that are close in the image space, are more likely to belong to the same class than pixels that are widely separated. This paper presents a mechanism that allows classifier learning to be reinforced by constraints such as spatial correlation. This involves reinforcement of classifier labelling decisions that satisfy the constraints. In comparison with clustering methods it offers a computationally more efficient scheme and better boundary localisation.