Statistical Modeling Of Global Structure For Handwritten Chinese Character Recognition

Handwritten Chinese character recognition is an important topic in pattern recognition and artificial intelligence. In all methods for handwritten Chinese character recognition, Statistical-structural methods receive extensive attention. This thesis focuses on statistical-structural methods and its application to handwritten Chinese character recognition, including stroke-point matching and its extraction, statistical modeling of character structure.An approach to stroke-point matching and extraction is proposed. We uniformly sample points in each stroke of reference character, and then extract the input character's feature point. To be more precise, we match the strokes firstly, in which stroke similarity measure is given. Minimum risk algorithm is improved, taking into account the maximum similarity criterion, which determines the stroke correspondence. Due to stroke connection, several-for-one mapping exits. We propose sorting algorithm for connective stroke, finally matching the strokes. Here, we can construct two point sequences for connective stroke. Then, we apply the dynamic programming to the feature stroke-point matching. Considering the differences of reference character and input character, the point matching result is not good enough. Thus affine transformation is adopted to minimize the deformation. Feature matching and affine transformation are optimized alternately, and ultimately feature stroke-point is extracted. Experimental results on CASIA-OLHWDB1 datasets show that the proposed method is effective.A statistical modeling method of global structure is proposed for handwritten Chinese character recognition. The character is globally represented as a feature vector which is composed of local features at extracted stroke-points with the reference character's stroke order. The local feature at a point is described by its coordinates and tangent slope angle. We define the shape change between corresponding stroke-points with Mixed Gauss Model (GMM). Expectation Maximization (EM) is adopted to learn the initiative GMM's parameter, and then Max-Min posterior Pseudo-probabilities (MMP) is applied to learn the final GMM's parameter. In the recognition, the corresponding posterior pseudo-probability and the complicated discrepancy are computed for the input character and the reference character, determining the input character's class. We differentiate characters with similar forms from CASIA-OLHWDB1 datasets, and achieve good recognition rate, which proves our method is feasible and promising.