Using a wavelet-based descriptor to extract information on object shape

From LRDE

Abstract

The text-localization system developed by the laboratory currently uses a wavelet-based descriptor to separate text from background in natural images. Benchmarks show that this descriptor is fast and accurate when performing this task. Our main objective is to extract shape information when performing classification using the same descriptor as the one used to separate text from background. Such data is useful improving the accuracy of other steps in the text localization system. Moreover, because the descriptor used to extract shape information has already been computedthis approach is not CPU consumin. More than shape information, the wavelet based descriptor can be extended again to create a simple O.C.R. (Optical Character Recognition). However in order to extract this data without changing the structure of the descriptor, some changes must be done in the classification process and in the learning process. Various methods to achieve these goals are presented and compared in order to see what kind of information could be extracted with the wavelet-based descriptor.