Skip to Main Content
This paper proposes a powerful method that realizes image reconstruction from a feature space in optical character recognition. Due to the invisibility of a high dimensional feature space, it is difficult to fully understand advantages and disadvantages of the given feature space and search for more robust features. The proposed method consists of two parts. The first part is 2D shape morphing based on a mesh model via bilinear transformation. The second part is use of genetic algorithms for determining optimal morphing parameters. Given an arbitrary feature vector in a feature space the proposed method deforms each category's template to yield the maximal fitness value against the given feature vector and the deformed template thus obtained is considered as a reconstructed image from a feature space. In experiments we use the public handwritten numeral database IPTP CDROMIB and a gradient feature space. We first demonstrate a high matching ability of the proposed mesh model. Then, we show promising experimental results of image reconstruction from a feature space and discuss how to use this technique to improve recognition performance.