View article

The extraction of distinctive image features is the most important step in pattern recognition systems due to their direct impact on learning the machine commonly used in these types of systems. In this paper, we propose a handcraft feature learning, which based on local distinctive image descriptors, for multispectral palmprint representation and recognition. In the training phase, a projection matrix (hash functions) and a codebook are obtained using the Pixel Difference Vectors (PDVs) of non-overlapping sub-blocks, in order to use it as prior knowledge in the feature extraction step. For the test phase, the extracted PDVs are encoded into binary codes using the projection matrix, then pooled as a histogram feature using the codebook. The experimental results carried out on the CASIA database show that the proposed framework achieves better performances compared to the state-of-the-art methods, in …