In offline handwritten Chinese character recognition

the high variability of the handwriting strokes is the main cause for lowering the recognition performance

thus decreasing the variability of the handwriting strokes is one effective and important way to improve the recognition accuracy. To solve this problem

we propose a new method of optimal sampling features

which are developed from the prevalently used directional features by following procedures. Firstly

four directional factor images are generated from an input binary character image. Next

these four images are transferred through a low pass filter

and then these four low passed images are sampled. The image values at these sampling positions produce a feature vector that is defined as sampling features. In the case of the sampling positions are uniform and fixed

the sampling features are subject to stroke variations

and these stroke variations will increase the within class pattern variability. In order to compensate for stroke variations

the sampling positions should be adaptable to these stroke variations. That is

the sampling positions should be displaced against reference patterns to decrease the within class variability

on the other hand the smoothness of the displacement should be preserved to keep the character's primary structure unchanged. The sampling features satisfying above conditions are defined as optimal sampling features. These two conditions could be expressed as a constrained minimization problem

thus optimal sampling features could be solved in an iteration procedure. For the sake of saving the time cost

a coarse to fine strategy is utilized. Finally

optimal sampling features are obtained

the discrimination of features is increased; and the recognition performance is improved. In order to demonstrate the effectiveness of optimal sampling features

we apply it to the THCHR database and compare it with directional features. The result shows that sampling features achieve higher recognition accuracy than directional features.