Research on flexible image matching of two dimensional object

点击翻译按钮获取中文摘要

1. A new, affine-invariant image matching technique via KL (Kullback-Leibler) divergence minimization was proposed. First, we represented a grayscale image as a probability distribution. Second, we introduced KL divergence between an affine-transformed input image and a template using their corresponding probabilty distributions. Finally, we determined optimal affine parameters that minimize KL divergence using an iterative method.2. A drastic acceleration of GAT correlation to realize distortion-tolerant image matching with far less computational burden was proposed. The key ideas were separation of variables and generation of lookup tables in the original GAT computational model. A powerful combination of k-NN classification and accelerated GAT correlation techniques achieved the state-of-the-art recognition accuracy in handwritten numeral recognition.