Object segmentation using an array of interconnected neural networks with local receptive fields

Neural networks (NNs), such as multi layer perceptrons and radial basis function architectures, proved to be powerful tools in many problems where the objective is robust classification. However, in applications that require simultaneous segmentation and recognition, such as speech and handwriting recognition, NNs were used with much less success. In this work, we introduce an architecture for object segmentation/recognition that overcomes some limitations of classical NNs by utilizing contextual information. An important characteristic of our model is that recognition is treated as a process of discovering a pattern rather than a onetime comparison between a pattern and a stored template. Our network implements some properties of human perception and during the recognition emulates the process of saccadic eye movements. We contrast our model to hidden Markov models in application to segmentation/recognition of handwriting and demonstrate a number of advantages.