Bidimensional empirical mode decomposition-based unlighting for face recognition Academic Article in Scopus uri icon

abstract

  • © 2014 IEEE. A face recognition system must be capable of handling facial data with head pose variations or different illumination conditions. However, as these conditions are uncontrolled the requirement of better algorithms has become essential. We propose a Bidimensional Empirical Mode Decomposition-based unlighting method that preprocesses the luminance and the reflectance parts of an image. First, three luminance components are estimated using Bidimensional Intrinsic Mode Functions residuals. Second, a shadow removal procedure using recursive Retinex is applied. Third, the reflectance part is denoised using mean-Gaussian filters. After that, a new image is created multiplying each shadow-free luminance by the reflectance. The final output is obtained using the geometric mean on the newly acquired images. This algorithm has been tested in two 3D- 2D face recognition databases: UHDB11 and FRGCv2.0. The performance of BEMDU demonstrates an improvement of up to 15.42% when compared with the AELM, LBEMD, PittPatt, the baseline, and EA algorithms.
  • © 2014 IEEE.A face recognition system must be capable of handling facial data with head pose variations or different illumination conditions. However, as these conditions are uncontrolled the requirement of better algorithms has become essential. We propose a Bidimensional Empirical Mode Decomposition-based unlighting method that preprocesses the luminance and the reflectance parts of an image. First, three luminance components are estimated using Bidimensional Intrinsic Mode Functions residuals. Second, a shadow removal procedure using recursive Retinex is applied. Third, the reflectance part is denoised using mean-Gaussian filters. After that, a new image is created multiplying each shadow-free luminance by the reflectance. The final output is obtained using the geometric mean on the newly acquired images. This algorithm has been tested in two 3D- 2D face recognition databases: UHDB11 and FRGCv2.0. The performance of BEMDU demonstrates an improvement of up to 15.42% when compared with the AELM, LBEMD, PittPatt, the baseline, and EA algorithms.

publication date

  • April 10, 2015
  • April 10, 2015