Face image illumination quality assessment for surveillance video using KPLSR

Illumination condition is one of the most important factors that affect the face recognition performance. Face image illumination quality assessment can predict the face recognition performance under various illumination conditions, which will improve the accuracy and efficiency of the face recognition system. However, the quality scores calculated by the existing methods are weakly correlated with the performance of face recognition. Face images with high scores often present bad recognition performance, while face images with low scores unexpectedly present good recognition performance. To predict the recognition performance more accurately, a kernel partial least squares regression (KPLSR) based face image illumination quality assessment method for surveillance video is proposed in this paper. The mapping relationship between illumination conditions and face recognition performance is modeled using KPLSR. Taken the fact that different regions of face image have different importance on face recognition, the features of luminance and contrast of sub-blocks are extracted to reflect the illumination conditions. Matching score of the face recognition system is calculated to measure the recognition performance. Experimental results show that, compared with the existing methods, the quality scores calculated by the proposed method are strong correlated with the recognition performance. The proposed method can also meet the requirements of real time processing.