Feature-locked loop and its application to image databases

We present a new dynamical system called the “feature-locked loop”. The inputs to this feedback neural network are a set of feature vectors and a one-parameter function that characterizes the data. We show that the feature-locked loop is locally stable for one example of the characteristic function and determines the value of its unknown parameter. We apply this property of the feature-locked loop to the problem of sorting textures by their similarity. We use the feature-locked loop and a priori information to quantify the degree of similarity between the input image and the reported set of image as a whole. The prior knowledge is encoded in the form of the one-parameter function and a general assumption about the number of perceptual outliers in the reported set. The unknown parameter, computed by the feature-locked loop, is then related to the entire set of image features produced by the retrieval