Multi-scale discriminant representation for generic palmprint recognition

Palmprint shows great potential in biometric-based security due to its advantages of great stability, easy collection, and high accuracy. However, with insufficient training samples, how to extract discriminative features applicable in various scenarios is still challenging. This paper proposes a general palmprint recognition framework. First, based on the integrated Gabor filter (IGF), a multi-scale integrated Gabor convolutional network (MS-IGCN) with large and small receptive fields is built. Compared with the existing DCNN, MS-IGCN applies fewer trainable parameters to targetable grasp principle lines and wrinkles features. Notably, our constructed IGF in MS-IGCN completes the feature extraction. The fixed Gabor filters in IGF extract spatial features of different orientations, and the learnable filters learn complementary features that cannot be extracted by the fixed Gabor filters. Then, the multi-scale discriminant representation (MSDR) of palmprint is learned through MS-IGCN. Finally, to make MSDR flexibly applicable to various scenarios, the ProCRC is implemented. A large number of experiments on public databases are conducted. The experimental results show that compared with various advanced palmprint recognition methods, MSDR has the best recognition effect. In addition, we complete the palmprint collection in an unconstrained environment and established an unconstrained palmprint dataset. The experimental results on the unconstrained palmprint dataset fully demonstrate that MSDR maintains the best performance in practical applications.