Some properties of the Competitive Layer Model with application to object regions extraction

It is known that the competitive layer mode (CLM) implemented by Lotka-Volterra recurrent neural networks (LV RNNs) can be used for feature binding. A group of features with similar property can be bound into same layer, however, it is not known which layer a group can be bound to. This is a drawback in some practical applications since it may be required to know which layer a group of features can be bound to. In addition, while using the CLM of LV RNNs for large data set clustering, it is difficult to set appropriate parameters of the network to achieve good clustering results. In this paper, a method called dividing and fixing group method is proposed to overcome this two problems. This method contains two steps. In the first step, it divides a large data set into several small sub data sets with overlapping among neighborhood sub data sets. In the second step, the CLM of LV RNNs is applied to each sub data sets, all features in one group can be bound to same layer by initializing the value of neurons for overlap elements in processing sub data set with the final value of neurons for same elements in processed sub data sets. As one application of this method, it is used to extract object regions in some images.