Transformations and distortions tolerant recognition of numerals using neural networks
Rajeshwar Prasad Srivastava · 1991
This paper presents a multilayered artificial neural network model to recognize numerals independent of their sizes, positions, and orientations.We used a modified backpropagation algorithm to train the network.We preprocessed the numerals for "feature" extraction by calculating their moments.The moments are invariant under translation, scaling, and rotation transformations.We used the moments as input to the network rather than the digitized pattern itself.The network was able to recognize transformed and slightly deformed numerals.Parallel computational capability of the network makes it an attractive alternative for real-time commercial applications. INTRODUCCrlONThe recognition of numerals has several practical applications including optical page readers, and mail sorters.In order to provide a practical system the recognition algorithm should be tolerant of small distortions and possible transformations.Developing such algorithm is a difficult problem in designing pattern-recognition systems.Many authors [ 1,2,3,5,9,16 ] have tried to solve this problem but no satisfactory general theory exists.Often, the methods are based on extracting "features" which are invariant under transformations.The main difficulty in using the conventional techniques is that they are not fault tolerant.Neural networks have been found to be fault tolerant in pattern recognition [ 3,4,5,6,7,8,9,12 ].Fukushima and Miyaki [3] used a multilayered neural network, called "neocognitron," to recognize numerals.Their method is tolerant of small distortions and shifts in position but it is quite difficult to train a "neocognitron" to include rotational invariance as well.