Abstract:

A novel method is proposed for infrared face recognition in this paper. Firstly, thermal images are converted into modified blood perfusion images by a modified blood perfusion model to obtain consistent facial images without the effect of the ambient variations. Secondly, the modified blood perfusion data are decomposed into a wavelet packet tree using three scales’ discrete wave packet transform. Then, proper numbers of nodes with the highest recognization score are further transformed via discrete cosine transform (DCT). The features extracted from the reserved nodes in the DCT domain are used for recognition using the 3nn classifier in the Euclidean distance to obtain the recognition results. The results are finally fused to output the final recognition results. The experiments conducted illustrate that the  proposed method has better performance. While the recognition rate does not decrease based on the modified blood model compared to the blood model obviously and have even lightly improved in some cases, and more efficient features are extracted using the coefficient of DCT in the wavelet packet domain and decision fusion, which can improve the infrared face recognition performance.

Key words: infrared face recognition；modified blood perfusion model；wavelet packet;discrete cosine transform;decision fusion