Genetic improvements in illumination compensation by the discrete cosine transform and local normalization for face recognition

Face detection and recognition depend strongly on illumination conditions. In this paper, we present improvements in two illumination compensation methods for face recognition. Using genetic algorithms (GA) we select parameters of the Discrete Cosine Transform (DCT) and Local Normalization (LN) methods to improve face recognition. In the DCT method all low frequency components within an isosceles triangle, of side Ddis, are eliminated. The best results were reported for Ddis=20. In the LN method it is proposed to normalize the value within a window by the mean and standard deviation. Best results were reported for window sizes of 7x7. In the case of the DCT method, we assigned weights to eliminate the coefficients of the low frequency components using a GA. In the case of the LN method for a fixed window size of 7x7, we selected the normalization method by a GA. We compare results of our proposed method to those with no illumination compensation and to those previously published for DCT and LN methods. We use three internationally available face databases Yale B, CMU PIE and FERET where the first two contain face images with significant changes in illumination conditions. We used Yale B for training and CMU PIE and FERET for testing. Our results show significant improvements in face recognition in the testing database. Our method performs similarly or slightly better than DCT or LN methods in images with non-homogeneous illumination and much better than DCT or LN in images with homogeneous illumination.