Face recognition and clustering for home photos

In this work, we focus on clustering faces in home photos by face recognition technologies. We propose two methods to improve the approach based on a well-known algorithm, local binary patterns. The adoption of the partial matching metric improves the recognition accuracy under face pose variations, while the adoption of the Gabor filter improves the accuracy under noises and various illuminations. We evaluate our methods on two home photo sets. In both evaluations, the results show that our methods improve the performance in accuracy. Compared to baseline LBP methods, in both evaluations, the results show that our methods improve the performance in accuracy from 90.4% to 99.5% and from 94.7% to 99.6% in two home-photo data sets respectively. Even compared to Google Picasa, the number of clusters where there is only one photo (thus can not be merged with other clusters, also means not good), our methods show that the number of "single" clusters reduced by half can be achieved. Our experience also shows that GPU speedup for Gabor filter can reach 140 times, and the overall system plus clustering can thus have 10 times speedup for face recognition.