Adaptive LPQ: An efficient descriptor for blurred face recognition

Local Phase Quantization (LPQ) is a state-of-the-art blur-insensitive texture descriptor. The theoretical and empirical results show that the major energy point of the blurred images depends heavily on the blur type and level, but classical LPQ samples the local patch at predefined frequencies. In this paper, we extend LPQ to Adaptive LPQ (ALPQ) by adaptively setting the sampling frequency for various types of quantized blur kernels, where subspace-based Point Spread Function (PSF) Inference is applied to estimate the blur kernels for the test images. Experimental results on the FERET database (with artificially blurred) and the FRGC database (with real blurred) demonstrate that sampling the local patch at adaptive frequency could largely improve the face recognition performance of LPQ. Moreover, the recognition performance of the proposed ALPQ method is comparable to the state-of-the-art deblurring based methods, such as FADEIN+LPQ.