The Study Of Temperature Normalization Of Infrared Face Images Based On Transform Domain

Face recognition is the research focus of the field of computer vision and biometrics identification. Variant illumination and expression can make a more serious effect on face recognition. In order to overcome these problems, researchers use the infrared face images to make identification. The biggest advantage of the infrared face is not dependent on illumination, which makes it is possible to using infrared face recognition to make up the shortcomings existing in visible light face recognition.However, infrared face recognition still has its limitations. Infrared face recognition also suffers from the effect of ambient temperature as visible face recognition suffers from the effect of illumination. Aiming at improving the performance of infrared face recognition system, this paper make a research on temperature normalization of the infrared face using the idea the methods dealing with face illumination normalization and proposes two kinds of temperature normalization based on transform domainThe temperature normalization method of infrared face images based on Fourier transform domain can be described as the point that the testing infrared face images can be normalized by finding and replacing the feature points which suffer from serious effect from ambient temperature. Finally, the more robust feature can be extracted from normalized testing infrared face images for recognition. Our experiments demonstrate show that this temperature normalization method can effectively improve the performance of infrared face recognition system.The normalization method based on the dual tree complex wavelet transform domain can require the stable face features for recognition by discarding low and high frequency components of the subbands coefficients which suffer from serious effects from the variant ambient temperature. Firstly, infrared face images collected under different ambient temperature can be transformed to the dual tree complex wavelet domain. Furthermore, according to the statistical characteristics of the low and high frequency components, the low-frequency or high frequency components sensitive to variant ambient temperature can be found and set to zero. Then, the refined low and high-frequency coefficients can be used to reconstructe the testing infrared face images and the temperature normalization of testing sample can be implemented. Finally, we confirmed experimentally that this method is feasible, robust. Meanwhile, two important conclusions can be drawn:the low frequency subband coefficients suffer from less effect from ambient temperature than the high frequency subband coefficients in infrared face images; six high-frequency subbands suffer from different effects from the ambient temperature at decomposing level.