Race Classification Based On Periocular Features Fusion

Race classification is a process of recognizing the race of people according to the race-related information extracted from their images. With the continuous development of science and technology, the exchange and integration of different races become more and more common as the geographical isolation, which ever resulted in the differences of races during the evolutionary process, is disappearing. Meanwhile, the conflicts about history and culture among different races are also becoming frequent, and even many terrorist events are associated with the race issue. In this context, race classification based on face image seems to be more and more important in the area of artificial intelligence. Currently, the methods about race classification are generally based on the entire region of the face or periocular region or only a local region. Although these methods have achieved some results, the results seem to be not very good and these methods require commonly high quality images to ensure the stability of the recognition rate. In this paper, after an intensive investigation on the race-related information in the periocular region, a two-type race classification method based on the fusion feature of 5 local features is proposed. This proposed method not only achieves a higher recognition rate, but also have a weaker requirement on the quality of images. But this method still has certain limitation that it can only be applied to the classification of East Asians and Caucasians at present.Firstly, after the investigation of race-related information in the periocular region, 4 local features are proposed: the feature of upper eyelid concave pentagon region, the feature of inner eye region, the comparison feature between regions in the middle of eyebrows and above inner corner, and the distance feature between the upper eyelid and eyebrows. Then regions of these features are located, and suitable feature descriptors are selected to represent these features. In addition, these features are compared between East Asians and Caucasians.Secondly, the feature extraction and representation of iris region which is commonly used in the area of race classification is achieved and features are also compared between East Asians and Caucasians.Thirdly, the multi-feature fusion method based on the 5 local features in periocular region for the classification of East Asians and Caucasians is proposed. For a test image, its 5 local features are extracted and then these features are used by the KNN classifier to get 5 temporary results. At last, these results are used with the no weighted voting which is one of the classifier ensemble algorithms to determine the final result.Finally, the experiment is conducted on the OFD-FERET database and proves that the proposed method in this paper has greater classification accuracy. Meanwhile, two additional experiments are conducted to verify the stability of this method when images with glasses exist in the database and the size of images are variant.