| 摘要 |
A system and method for performing pattern matching to locate zero or more instances of a template image in a target image. The method first comprises sampling the template image using a Low Discrepancy sequence, also referred to as a quasi-random sequence, to determine a plurality of sample pixels in the template image which accurately characterize the template image. The Low Discrepancy sequence is designed to produce sample points which maximally avoid each other. After the template image is sampled or characterized, the method then performs pattern matching using the sample pixels and the target image to determine zero or more locations of the template image in the target image. The method may also perform a local stability analysis around at least a subset of the sample pixels to determine a lesser third number of sample pixels which have a desired degree of stability, and then perform pattern matching using the third plurality of sample pixels. In one embodiment, the local stability analysis determines a plurality of sets of sample pixels with differing stability neighborhood sizes, and the pattern matching performs a plurality of iterations of pattern matching using different sets of sample pixels, preferably performed in a coarse to fine manner, e.g., using sets of sample pixels with successively smaller stability neighborhood sizes and/or step sizes. The present invention also includes performing rotation invariant pattern matching by sampling the template image along one or more rotationally invariant paths, preferably circular perimeters, to produce one or more sets of sample pixels. These sample pixels from the circular paths are then used in the pattern matching. The rotationally invariant pattern matching may also use local stability analysis and coarse to fine searching techniques.
|
