Affine-transformation and 2D-projection invariant k-NN classification of handwritten characters via a new matching measure

• We enhance k-NN classification by affine and 2D-projection invariant matchings.
• We develop acceleration methods for those distortion-tolerant matching techniques.
• We propose a matching measure using similarity in direction and curvature of edges.
• Experiments using the MNIST database show a very low error rate of 0.30%.
• The source code used in the above-mentioned experiments is uploaded.

Pattern recognition based on matching remains important because it is a fundamental technique, it does not require a learning process, and the result of matching provides intuitive and geometrical information. Wakahara et al. proposed global affine transformation (GAT) correlation matching, which can compensate for affine transformations imposed on a pattern. GAT correlation matching with an acceleration method and a new matching measure, called the nearest-neighbor distance of equi-gradient direction (NNDEGD), achieved high performance in experiments using the MNIST database. The GAT matching measure was extended to a global projection transformation (GPT) matching measure to allow deformation by 2D projection transformations.The purpose of this paper is threefold. First, we develop an acceleration method for GPT correlation matching. Second, in order to improve recognition performance, we apply the curvature of edges in strokes to the matching measure. Curvature is often used as a feature of characters. However, in this paper, we use it as a weight in the NNDEGD. Third, to investigate the performance of the proposed methods, we apply them to image matching and recognition from the MNIST and the IPTP databases for k-nearest neighbors (k-NN). In the experiment with the MNIST database, the GPT correlation matching with the curvature-weighted NNDEGD matching measure achieves the lowest error rate of 0.30% among k-NN based methods.