| 摘要 |
The disclosed method comprises a learning phase, during which the object zone to be recognized, selected in a learning image, is modelized, and a recognition phase during which the modelized zone is compared with image zones of a same size taken taken from the images to be processed. To characterize the image zones that are models of objects to be recognized or image zones to be processed, the method uses a decomposition of the matrices of the luminance values into singular values according to a matrix product of projection matrices and an ordered diagonal matrix, this ordered diagonal matrix constituting a characteristic signature of the zone considered; for the recognition phase, differences between the signature of the model zone and the signatures of the zones of the image considered are computed and image zones are selected on the basis of the least differences according to an iterative method which consists in conducting a rough sorting operation on a large number of zones in an initial stage, and then an increasingly finer sorting operation on an increasingly limited number of zones. The computations done in each step are re-used in the next step.
|
