2
different position. So a specialized pen is necessary for handwriting on the piezoresistive
touch panel. Unlike the piezoresistive touch panel, the capacitive touch panel enables the
user to handwrite directly on the panel without any intermediate device. But the capacitive
touch panel cannot be used under wet conditions due to the sensitivity to the distributed
capacitance of the capacitive touch panel. Recently, Wu [7] proposed a temperature sensing
terminal, which could be used as a new type handwriting inputting terminal, and it was not
sensitive to changes of the pressure, the distributed capacitance, and the environment
humidity. So it was an attractive complement to traditional HCI interfaces. We addressed
an online handwritten digit recognizer for the temperature sensing terminal in this paper. In
the temperature sensing terminal, the sensitive panel had an 8×16 temperature sensor array,
which was robust against the variation in the environment. But the high real time
requirement of the online handwritten digit recognition limited the spatial resolution of the
handwriting inputting information, which brought difficulties to the online handwritten
digit recognition.
The automatic handwritten digit recognition normally includes three steps: preprocessing,
feature extraction, and classification. The performance of handwritten digit recognition
largely depends on the feature extraction approach and the classification scheme. The
feature extraction is one of the crucial steps in the recognition system. The common
features include static features and dynamic features [8-10]. Static features are insensitive
to the writing sequence of the digit, but are susceptible to local noise. Dynamic features are
not sensitive to noise and small deformation of handwritten digits, but the different stroke
order of digits will cause wrong recognitions. To enhance the judgment of digit features, the
hybrid features including static features and dynamic features are extracted. Classification
is another crucial step in the recognition system. To improve the performance of
handwritten digit recognition systems, many classification methods have been proposed,
such as statistical methods [11-17], rule inference methods [18, 19], neural network
methods [20-25], and so on. Statistical methods such as Bayesian method [11], Fisher
method [12, 13], k nearest neighbor (kNN) method [14, 15], and support vector machine
(SVM) method [16, 17], are based on the assumption of the probability density. The
disadvantage of the statistical methods is that the small probability event has been ignored
[11]. The decision tree method [18] and the association rules method [19] are rule inference
methods. The rule inference methods can make full use of the structure characteristics of
the handwritten digits. But they demand a lot of expert knowledge, which requires much
manual work. The neural network can store information in distribution, process
nonlinearity, and tolerate much fault. So many kinds of neural networks are used to
handwritten digit recognition. The multilayer perceptron (MLP) [20, 21] is the first neural
network classifier tested on MNIST and it is the simplest neural network classifier used to
handwritten digit recognition. The back-propagation neural network (BPNN) classifier [22,
23] is overall superior in memory usage and classification time but it may provide “false
positive” classifications when the input is not a digit. The radial basis function neural
network (RBFNN) classifier [24] can provide the similar low error rate with the kNN
classifiers. The RBFNN classifier requires more memory and more classification time. The
number of hidden units of the neural network has a significant impact on the performance
of the neural network classifier, but it is difficult to be determined. The cascade-correlation