Feature Vector Compression based on Least Error Quantization

We propose a distinctive feature vector compression method based on least error quantization. This method can be applied to several biometrics methods using feature vectors, and allows us to significantly reduce the memory size of feature vectors without degrading the recognition performance. In this paper, we prove that minimizing quantization error between the compressed and original vectors is most effective to control the performance in face recognition. A conventional method uses non-uniform quantizer which minimizes the quantization error in terms of L-2-distance. However, face recognition methods often use metrics other than L-2-distance. Our method can calculate the quantized vectors in arbitrary metrics such as L-p-distance (0 < p <= infinity) and the quantized subspace basis. Furthermore, we also propose a fast algorithm calculating L-p-distances between two quantized vectors without decoding them. We evaluate the performance of our method on FERET, LFW and large face datasets with LBP (L-p-distance), Mutual Subspace Method and deep feature. The results show that the recognition rate using the quantized feature vectors is as accurate as that of the method using the original vectors even though the memory size of the vectors is reduced to 1/5 - 1/10. In particular, applying our method to the state-of-the-art feature, we are able to obtain the high performance feature whose size is very small.