Alexander Shustorovich

A subspace projection approach to feature extraction: the two-dimensional Gabor transform for character recognition

Abstract This paper describes an application of the two-dimensional Gabor wavelets as feature extractors for character recognition with neural networks. Our approach is based on an analysis of the function performed by a single hidden unit in the first layer of a network presented with raw pixel data. This weight function can be approximated by a linear combination of basis functions from a fixed set. We establish the duality between this expansion and feature extraction: the projections of an image onto the same basis set play the role of precalculated features, and they are used as the input to the network. Recognizability of images reconstructed from these projections suggests that the necessary information is preserved by the corresponding feature extraction scheme. In this study, the Gabor wavelets provided the best trade-off between dimensionality reduction and quality of the reconstructed images. A local receptive field (LRF) network was trained on the NIST data base of isolated alphanumeric characters and tested on unseen parts of the same data base. The use of Gabor projections instead of original pixel data resulted in improvement from 86.35% to 89.40% for the lowercase, from 89.40% to 96.44% for the uppercase, and from 98.63% to 99.11% for digits, which corresponds to 22–66% reduction of classification error. This LRF-Gabor network became a part of a unified algorithm used by Eastman Kodak Company that finished in the tight group of leaders at the U.S. Census Bureau/NIST First OCR Systems Competition.

Neural network positioning and classification of handwritten characters

Abstract This paper describes two algorithms at the core of the new Kodak Imagelink™ OCR numeric and alphanumeric handprint modules. Both variants of the system were designed to work with fields of characters, typically scanned from forms. The first neural network is trained to find individual characters in the field. Its outputs are associated with an array of pixels in the middle of a sliding window, and they signal the presence of characters centered at corresponding positions. A window containing each detected character (and, possibly, pieces of adjacent characters) is passed on to the second network, which performs the classification. The outputs of both networks are interpreted by an application specific postprocessing module that generates the final label string. Both networks were trained on Gabor projections of the original pixel images, which resulted in higher recognition rates and greater noise immunity. The system has been implemented in specialized parallel hardware, and has been installed and used in production mode at the Driver and Vehicle Licensing Agency (DVLA) in the United Kingdom. The success rate of the purely numeric handprint module (as measured on randomly selected batches of over 200 real forms containing 3500 characters) exceeds 98.5% (character level without rejects), which translates into 93% field rate. After approximately 7% of the characters are rejected, the system achieves a 99.5% character level success rate acceptable for this application. The similarly measured overall success rate of the alphanumeric handprint module exceeds 96% (character level without rejects), which translates into 85% field rate. If approximately 20% of the fields are rejected, the system achieves 99.8% character and 99.5% field success rate.

Scale specific and robust edge/line encoding with linear combinations of Gabor wavelets

Abstract A method of detection of characteristic orientations of local image structure at a specific scale of analysis and at a specific location is described. The technique is almost immune to noise, and it results in a structural description that can be used by higher level image analysis and pattern recognition algorithms. The approach is based on rather unusual properties of two-dimensional Gabor wavelets, namely their ability to model rotated, scaled, and shifted versions of themselves with linear combinations of a discrete basis set. The paper is concluded with the results of a character recognition experiment, in which a simple template-matching procedure based on the structural description achieved 99.5% correct classification on a test from the NIST database of pre-segmented digits.

Reducing the Weight Space of a Net With Hidden Units to a Minimum Cone.

In his recent talk [1] on the theory of Back-propagation (BP) at IJCNN-89, Dr. Hecht-Nielsen made an important observation that any single meaningful combination of weights can be represented in the net in a huge number of variants due to the permutations of hidden units. He remarked that if it were possible to find a cone in the weight space such that the whole space is produced from this cone by permutations of axes corresponding to the permutations of the hidden units, it would greatly reduce the volume of space in which we have to organize the search for the solutions.