MIultiscale Object Recognition and Feature Extraction Using Wavelet Networks (U)*

In this work. we present a novel method of object recognition and feature generation based on multiscale object descriptions obtained using wavelet networks in combination with morphological filtering. First, morphological filtering techniques are used to obtain structural information about the object. Then, wavelet networks are used to extract or capture geometric information about an object at a series of scales. A wavelet network is of the form of a 1-1/2 layer neural network with the sigmoid functions replaced by wavelet functions. Like neural networks, wavelet networks are universal approximators. In contrast to neural networks. the initialization of a wavelet network follows directly from a commonly known transformation. namely, the discrete dyadic wavelet decomposition. In contrast to a dyadic wavelet decomposition, the wavelet parameters are then allowed to vary to fit the data. Although developed in the context of function approximation. wavelet networks naturally fit in this object recognition framework because of the geometric nature of the network parameters (i.e. translations, rotations, and dilations). %Wavelet networks are the basis for a hierarchical object recognition scheme where the wavelet network representation of the object at each scale is a feature vector which may be used to classify the object. At coarse scales, the feature vector is used to narrow the field of possible objects and to yield pose information. This information may also be used to generate candidate matches between the data and more detailed object models. The wavelet network representation at finer scales is then used to identify the object from this reduced space of possible objects. In keeping with our proposed integrated approach to ATD/R, we demonstrate how wavelet networks may be applied to anomaly suppression in laser range images by fitting a multiresolution wavelet basis to the data in conjunction with the expectation-maximization (EM) algorithm. Here, the wavelet network is used to refine the chosen wavelet basis. \We demonstrate each of these applications on both simulated and model board data. *This work was supported by the Advanced Research Projects Agency through Air Force Grant F49620-93-1-0604 and the Army Research Office though Grant DAAL03-92-G-0115 and the Air Force Office of Scientific Research through Grant F49620-92-J-0002.