Boosting palmprint identification with gender information using DeepNet

Abstract

Abstract Palmprint has been found to possess distinct features for reliable person identification and gender classification. Traditionally, the palmprint-based recognition approaches often rely on hand-crafted feature descriptors, which do not migrate well to the emerging datasets. This is because the local textural features cannot always be well preserved under complex acquisition environment. In recent years, deep convolutional neural networks (CNNs) have greatly advanced the performance of biometric applications even encountering non-ideal images caused by motion blur, poor contrast or illumination artifacts. However, CNNs have not been fully explored in palmprint-based biometric recognition and gender classification, mainly owing to lack of diverse training data with gender labels. In this paper, our goal is to prove a CNN model that can be successfully used for palmprint identification and gender classification. Also, we tend to prove the hypothesis that gender information can help to boost palmprint identification performance. For this end, we first proposed two palmprint datasets: BJTU_PalmV1 and BJTU_PalmV2 with gender attributes, captured by CCD cameras and mobile phones, respectively. BJTU_PalmV1 contains 2431 hand images from 174 persons, whilst BJTU_PalmV2 contains 2663 hand images from 148 subjects. Secondly, two deep CNN-based models are separately designed to accomplish palmprint identification and gender classification in an end-to-end trainable fashion using palm images. Thirdly, we introduced two novel deep CNN architectures to validate the hypothesis that gender information is conductive to identification accuracy. The experimental results conducted on six different datasets: PolyU2D/3D1.0, IITD, Tongji, CASIA, BJTU_PalmV1 and BJTU_PalmV2, indicate that the proposed approaches achieve high accuracy and strong generalization ability. Also, the experimental results solidify that gender information can help in boosting palmprint identification.