The design of matching pursuit filters.

This paper presents a new technique for creating efficient and compact models from data, called matching pursuit filters. The design of a matching pursuit filter is based on an adapted wavelet expansion, where the expansion is adapted to both the data and the pattern recognition problem being addressed. This contrasts with most adaptation schemes, where the representation is a function of the data, but not of the problem to be solved. This approach does not decompose the images in the training set individually, but rather determines the expansion by simultaneously decomposing all the images. Because it uses two-dimensional wavelets as the building blocks for the decomposition, the representation is explicitly two-dimensional and is composed of local information. Matching pursuit filters can be trained to detect, recognize, or identify objects and have been applied to recognizing faces and detecting objects in infrared imagery.