Shape classification of partially occluded objects using subspace detectors

In this paper we consider shape classification of partially occluded objects. We model the occlusion as non-Gaussian noise, and apply robust subspace detectors in the classification module. We show that the robust subspace detectors can be formulated as a weighted subspace detector, and the elements in the boundary vector will be weighted before they are matched. The part of the boundary vector that corresponds to the occluded part of the object, will be suppressed by the weight vector, and hence have a reduced effect on the classification performance. The detectors are demonstrated on fish species classification applications.