Research On The Evolutionary Template Matching Algorithm And Its Applications

Current image matching technique suffers from the problem that high precision in matching normally incurs an expensive computational cost. To deal with the problem, in this dissertation, image matching is formulated as an optimization problem based on multi-peaks of the correlation between source image and template. To reduce the computational cost, a novel evolutionary template matching algorithm is proposed where a search strategy based on the evolutionary algorithm is employed. The main work of the dissertation are summarized as follows:A novel evolutionary template matching algorithm. Prior research work has shown that traditional correlation-based methods can acquire high matching precision and have strong flexibility in various matching problems. However, high computational complexity due to the pixel-by-pixel search strategy is one major drawback of the correlation-based methods. In this dissertation, a much faster image matching algorithm is developed. Reduction on the computational complexity is achieved by using an evolutionary algorithm based search strategy, due to the non-exhaustive search nature of the algorithm.Research on the theory of evolutionary template matching algorithm. There are two problems with the evolutionary template matching algorithm in practical applications. Firstly, the convergence rate is slow. Secondly, the algorithm is prone to prematurity. To speed up the convergence rate, this dissertation develops a competitive initialization method with controlled individual gap. The new method could guarantee the diversity of the initial population, and meanwhile make the initial seeds fall into the nearby areas of local optimums with greater probability, which helps to greatly reduce the searching time. To tackle with the second problem, this dissertation borrows the idea of simulated annealing and proposes an effective selection method-anneal selection strategy.Research on the applications of the evolutionary algorithm. By 2-dimensional wavelet decomposition technique, each face image can be decomposed into four subband images. After analyzing and evaluating the performance of the four subband images, the low frequency subband image is shown to be most suitable for facial features. Thus, we present a novel face detection method based on the evolutionary template matching algorithm. Different from most face detection techniques which can only detect the head-on faces, our method can quickly detectthe faces from images with varying positions and orientations.