| Face recognition has, drawn a lot attention and becomes research focus of computer vision, pattern recognition, machine learning and other areas, since it has overwhelming advantages against traditional authentication methods. However, since face images are sensitive to illumination and age, how to achieve a real-time and accurate face recognition algorithm is a real challenge.Even researchers have come up with many feature extraction algorithms; studies show that one feature alone is difficult to achieve the desired results. Thus, in this paper we propose a feature fusion method based on AdaBoost algorithm. The main work is as follows:1. We choose Gabor filter and POEM code images as our face image features. Describes Gabor filter related knowledge. The great majority of mammalian cortical simple cells have 2-D receptive filed profiles which can be well fit by members of the family of 2-D Gabor elementary functions. So the Gabor filter has a great advantage in the images feature extraction. We sample Gabor features using an existing grid sample method.2. We introduce a new face feature, POEM histogram. It uses the LBP operator on the gradient images. It’s robust to the illumination changes. For POEM code images, the patches with different size are traversed all over the image. We calculate the patch code histogram as features.3. We use the AdaBoost algorithm to fuse the Gabor and POEM features. The similarities of the two features between two images are arranged in a vector. Then AdaBoost is used to select the features. We choose different Gabor feature points and POEM histogram patches on different location. The features are given different weights, so that the size of features and the correlation between them are both reduced. We train the classifier on CAS-PEAL face dataset, and test our algorithm on FERET dataset. We achieve 99.0%,99.5%,87.8%,86.6% recognition rate respectively on Fb, Fc, Dupl, Dup2 sub set. It shows that, our feature fusion method is better than others’. |
PDF Full Text Request