Discriminant spectral hashing for compact palmprint representation

When palmprint recognition needs to be run in the device with low processing and small storage capacities, binary representation with low storage overhead, high matching speed and high discrimination power is preferred. However, existing feature extraction methods focus more on matching accuracy than representation compactness, which would result in high storage and operation cost. Inspired by Spectral Hashing that is known for compact-binary-representation extraction in the image retrieval domain, we propose a compact binary feature extraction method called Discriminant Spectral Hashing (DSH). DSH projects the feature to a discriminative subspace and then performs Spectral Hashing to obtain discriminative and compact code. Experiment results on a benchmark palmprint database show that our algorithm outperforms the existing coding-based methods in recognition accuracy with shorter code.