A Multi-Slice Encoding Direction Extraction Network for Palmprint Recognition

Palmprint recognition stands out for its exceptional accuracy and accessibility, substantially enhancing personal security in mobile payments and identity verification applications, as well as efficiently distinguishing individuals through the unique patterns of a person's hand. The distribution of viable palmprint textures across images is frequently uneven, but follows a recognizable pattern. Typically, the upper region of the palm shows a dense arrangement of lines that are predominantly horizontal, whereas the lower region is characterized by a sparser distribution of lines that are more vertically aligned. However, existing palmprint recognition methods have not yet thoroughly explored the differences in textural directionality among various local regions of palmprints. To address this issue, we propose an innovative Multi-Slice Encoded Direction Extraction Network (MSEDNet) that segments the palmprint images into multiple smaller slices for palmprint recognition. Specifically, we design a Learnable Direction Extraction Block (LDEB) to adjust the sensitivity between the X and Y directions independently for each slice. The texture density within a slice is closely related to the relative positions between slices and the central area usually displays a denser texture than the peripheral regions. Furthermore, we develop a dual-branch structure that incorporates both the Relative Position Encoding Branch (RPEB) and the Texture Density Encoding Branch (TDEB) integrating the relative positional information between slices with the texture density information within each slice. Extensive experimental results on five public palmprint datasets demonstrated that our proposed method achieves remarkable recognition results.