GAD: A Global-Aware Diversity-Based Template Matching Method

A novel similarity measurement metric, global-aware diversity (GAD), between a template and an image is proposed in this article, which can be efficiently utilized in a template matching method. Unlike the existing nearest neighbor field (NNF)-based methods, which use either the local or global diversity alone, GAD utilizes the global context to guide the local diversity. We verify that the global context is important for template matching task and can largely remove the background and outliers. The state-of-the-art NNF-based methods in general are not efficient enough as resolution grows. Thus, we also elaborately design a very efficient algorithm of GAD to reduce the complexity from O IT) to O (I). First, the GAD algorithm has been evaluated on two challenging benchmark datasets best-buddies similarity (BBS) and TinyTLP. Experiments show that the GAD algorithm achieves the state-of-the-art performance. As GAD does not impose any prior on unseen data, it is more insensitive to large rotation or deformation than the deformation-based algorithms. Besides, the GAD algorithm can run at 150 ms on average with high accuracy on the dataset of resolution 1280 ×780.