T.T. Son and S. Mita (Japan)
Face detection, Radon transform, Image-size pyramid, Multi subspace, Adaboost.
This paper presents an improvement of face detection in processing time. The proposed method utilizes the Radon transform to project an eye image from 2-D to multi 1-D data in radial directions. For each radial direction, a sub space is generated and saved into database in the learning process. In the test process, an input image is segmented into many blocks. Then each block is classified to be an eye region or not by matching its line integral to the eye databases in terms of multi subspaces of line integral. Af ter matching and classifying, a rectangle boundary of one confirmed-eye candidate is spanned and recorded as a can didate of face. Then, each face candidate is classified by the Adaboost method. Using this method, we can improve processing time by reducing the candidate of face consid erably. It can reduces about a half of processing time of the conventional Adaboost method. The proposed method reach 89% in the CMU database with 203 images and 436 faces.
Important Links:
Go Back
