An Accurate Extraction of Facial Meta-Information Using Selective Super Resolution from Crowd Images

초록

An accurate extraction scheme of facial meta-information from low resolution crowd images is proposed. In order to detect crowd abnormal situations, extracting facial meta-information from crowd images is very helpful. However, since many crow images are of low resolutions, extracting facial meta-information is pretty difficult. To extract accurately facial metainformation from the low resolution crowd images, some face images which are not suitable to easily extract the meta-information should be classified and be improved the quality by a super resolution method. To confirm the feasibility of the proposed scheme, since gender of person can be regarded as very important facial meta-information, we compare the gender classification accuracies of using the proposed scheme and that of using the input crowd image itself. According to the experiment results, the proposed facial extraction scheme from crowd images using selective super resolution can improve the gender classification accuracy for crowd images than that for using the original crow images.