Unconstrained Handwritten Digit Recognition Research Articles

Unconstrained handwritten digit recognition using perceptual shape primitives

In this paper, we propose a new handwritten digit recognition method which works in a very similar way as human perception. The digit image boundary is decomposed into four salient visual primitives, namely closure, smooth curve, protrusion and straight segment by defining a set of external symmetry axis. Unlike the conventional algorithms, our low complexity shape decomposition method neither searches for curvature minima nor finds optimal parsing by using shortcut and convexity rules. Based on the spatial configuration of extracted primitives, the recognizer classifies a test digit image using a set of classification rules. The performance of our proposed recognition system is evaluated on five digit datasets of four popular scripts, Odia, Bangla, Arabic and English. The recognition accuracies on the ISI Kolkata Odia and Bangla, IITBBS Odia, CMATERdb Arabic and MNIST English digit datasets are found to be 99.02, 99.25, 99.66, 97.96 and 99.11%, respectively. The proposed method outperforms the existing recognition systems on both the Odia digit datasets and achieves comparable performance in other cases.

Incremental class learning approach and its application to handwritten digit recognition

Incremental Class Learning (ICL) provides a feasible framework for the development of scalable learning systems. Instead of learning a complex problem at once, ICL focuses on learning subproblems incrementally, one at a time — using the results of prior learning for subsequent learning — and then combining the solutions in an appropriate manner. With respect to multi-class classification problems, the ICL approach presented in this paper can be summarized as follows. Initially the system focuses on one category. After it learns this category, it tries to identify a compact subset of features (nodes) in the hidden layers, that are crucial for the recognition of this category. The system then freezes these crucial nodes (features) by fixing their incoming weights. As a result, these features cannot be obliterated in subsequent learning. These frozen features are available during subsequent learning and can serve as parts of weight structures built to recognize other categories. As more categories are learned, the set of features gradually stabilizes and learning a new category requires less effort. Eventually, learning a new category may only involve combining existing features in an appropriate manner. The approach promotes the sharing of learned features among a number of categories and also alleviates the wellknown catastrophic interference problem. We present promising results of applying the ICL approach to the unconstrained Handwritten Digit Recognition problem, based on a spatio-temporal representation of patterns.

Aggregating membership values by a Choquet-fuzzy-integral based operator

A new type of aggregator, the Choquet-fuzzy-integral based operator, is introduced and applied in aggregation of membership values. In this paper, we generalized the definition of the fuzzy integral as a tool for membership aggregation. The generalization involves replacing the max (or min) operator in information integration with a Choquet-fuzzy-integral based aggregator, resulting in increased flexibility. The characteristics of the Choquet fuzzy integral are studied and a two-phase aggregating process is proposed. Also, we demonstrate a good classification result using the proposed aggregator to integrate the memberships of several clusters in a complex unconstrained handwritten digit recognition domain.

Recognition of handwritten digits based on contour information

In this paper, a new character recognition method based on topological properties, positions of starting point, statistical analysis, and shape recognition is proposed for recognition of unconstrained handwritten digits based on various contour information. First, input digits are classified into three groups according to their topological properties. In group 1, the features of normalized length of the digit contour, the normalized area of the digit and Fourier descriptors of the outer contour of a digit are used for recognition of digits. In group 2, the relative position of outer and interior contour centroids, the position of interior contour centroid, the mean and the variance of the outer contour distance function, and Fourier descriptors of the outer contour of a digit are used for the recognition of digits. In group 3, usually there is only one digit 8, but because of the complexity of writing styles, some other digits are still classified into this group. The final recognition is based on shape comparison of the input digit with models. In the recognition process, some special models are established and used for recognition of broken digits and digits whose topological properties are destroyed. In our experiment, 1000 digits from the NIST database are used for training and 5278 unseen digits are used for testing. The recognition rate has reached 98.5% with a reliability rate of 99.09%, a substitution rate of 0.91% and a rejection rate of 0.59%.

Energy function for learning invariance in multilayer perceptron

A new energy function is proposed for forming self-adapting ordered representations of input samples in a multilayer perceptron. Simulation results on unconstrained handwritten digit recognition give a better invariance extraction for this model than for several other models.

An Efficient Method to Construct a Radial Basis Function Neural Network Classifier

Radial basis function neural network (RBFN) has the power of the universal function approximation. But how to construct an RBFN to solve a given problem is usually not straightforward. This paper describes a method to construct an RBFN classifier efficiently and effectively. The method determines the middle layer neurons by a fast clustering algorithm and computes the optimal weights between the middle and the output layers statistically. We applied the proposed method to construct an RBFN classifier for an unconstrained handwritten digit recognition. The experiment showed that the method could construct an RBFN classifier quickly and the performance of the classifier was better than the best result previously reported.

Fourier descriptors and handwritten digit recognition

This paper presents the results of a comparative study of various Fourier descriptor representations and their use in recognition of unconstrained handwritten digits. Certain characteristics of five distinct Fourier descriptor representations of handwritten digits are discussed, and illustrations of ambiguous digit classes introduced by use of these Fourier descriptor representations are presented. It is concluded that Fourier descriptors are practically effective only within the framework of an intelligent system, capable of reasoning about digit hypotheses. We describe a hypothesisgenerating algorithm based on Fourier descriptors which allows a classifier to associate more than one digit class with each input. Such hypothesis-generating schemes can be very effective in systems employing multiple classifiers. We compare the performance of the five Fourier descriptor representations based on experiment results produced by a particular hypothesis-generating classifier for a test set of 14000 handwritten digits. It is found that some Fourier descriptor formulations are more successful than others for handwritten digit recognition.

The efficiency of holistic template matching in the recognition of unconstrained handwritten digits

Psychological evidence suggests that simple visual patterns can be recognized by the use of internal representations as holistic templates, but the efficiency of holistic template matching in the recognition of real-life patterns, such as handwritten characters, has been doubted. To clarify this issue, we measured the efficiency of holistic template matching in machine recognition of totally unconstrained handwritten digits. Our learning and recognition algorithm was simple; no previous knowledge of handwritten digits was presupposed, and preprocessing was limited to Gaussian smoothing and normalization with respect to position, size, and orientation. For patterns presented in a known orientation, recognition rates were .69, .77, and .88, respectively, when about 5, 10, or 50 templates had been learned for each type of digit. For patterns presented in unknown orientations, recognition rates were slightly lower. High levels of reliability could be attained by the discounting of classifications based on weak evidence. Apparently, in high reliability recognition, holistic template matching can be used as a first operation by which recognition is achieved for most of the handwritten digits that are seen in real life.