Image classification using expansion matching filters and hidden Markov trees

Images of three-dimensional objects are characterized by the different object components visible in the image, with the visible components being dependent on the object-sensor orientation. We therefore define a class as a set of contiguous object-sensor orientations over which the associated object image is relatively stationary with aspect, and therefore each object is in general characterized by multiple classes. Our feature parser employs a distinct set of expansion matching (EXM) filters for each class of images, to identify the presence of object subcomponents. Often, there is not simply one realization of the image of an object at a given orientation, but rather an ensemble of images accounting for variable imaging conditions. For example, in infrared images the visibility of a particular sub-component of an object and its intensity depends on the temperature of that sub-component, which in turn is influenced by the temperature of the neighboring subcomponents. The fundamental contribution in this research is the hidden-Markov-tree modeling of such variations among neighboring subcomponents of the images belonging to the same class.; The algorithmic performance demonstrated on a data set of forward looking infrared images (FLIR) of vehicles, is compared with that of other competing classifiers: wavelet-feature-based hidden Markov tree (wavelet-HMT), Radon-transform based hidden Markov model (RT-HMM) and expansion-matching feature based linear vector quantization (EXM-LVQ). The wavelet filters are not well suited for identifying the presence of different subcomponents in an image. The classification in the RT-HMM is largely based on the size of the object. In the EXM-LVQ algorithm, we use the same features as in the EXM-HMT algorithm, however the relation between neighboring subcomponents is characterized by learning vector quantization. We achieve a successful classification rate of 91% on a data set of infrared images of vehicles, compared with 64% for the wavelet-HMT algorithm. 70% for the RT-HMM, and 84% for EXM-LVQ algorithm. While the algorithm is demonstrated on infrared images, it is readily applicable to wide variety of image classification problems.