Research on Matrix Similarity Based on Two-Dimensional Dynamic Time Warping Algorithm

• 1、School of science,Jiangnan University, Wuxi 214122

Abstract：As a commonly used similarity measure in pattern recognition, distance is widely used in clustering and classification algorithms. With the extensive applications of matrixes, traditional algorithms have limitations in measuring similarity of matrixes which has more and more complex content. Aiming at this problem, the two-dimensional dynamic time warping algorithm (2D-DTW) is proposed. In 2D-DTW algorithm, Euclidean distance is used to construct the distance-cuboid between matrixes, and then the distance-matrix is defined by cutting and accumulating the distance-cuboid, finally, the traditional DTW algorithm is used to calculate the similarity between matrixes. By constructing distance-cuboid and introducing alignment mechanisms, 2D-DTW algorithm can improve the accuracy in calculating the distance between matrixes. The experimental result of CIFAR-10 database verifies that the 2D-DTW distance is superior to the traditional Euclidean distance in comparing the foreground target in the image. What\'s more, 2D-DTW algorithm can control the calculation time within a reasonable range. In addition, in the handwritten digit recognition experiment of the MNIST dataset, the recognition accuracy of the 2D-DTW algorithm is 21.50% higher than that of the traditional Euclidean distance. At the same time, the 2D-DTW algorithm works better in the physical fingerprint experiment.

Keywords： Pattern Recognition Similarity Distance Dynamic Programming DTW