Palmprint recognition as a novel biometric identification method for contactless mobile devices has been received substantial attentions in recent years. Palm landmark detection is one of the key technologies of palmprint identification and verification system. However, the differences
of hand positions, complex backgrounds and various lighting conditions in unrestrained environment with low-resolution cameras make palm landmark detection in the wild difficult. In this paper, we proposed a new palm landmark detection approach based on Supervised Descent Method (SDM). SDM
uses the relationship between the feature representation and the position of a landmark point to build an optimization problem for palm landmark detection. The optimization target function is the distance of feature representations between current position and the ideal position of a palm
landmark point. After optimization, a linear function of the position displacement and the feature representation of current landmark is obtained. The linear function can be learned from palmprint image samples with labeled landmark positions. Given an input image in detection process, the
initial position of a landmark is set by the mean position of the landmark in the training set, then the optimal landmark position can be calculated iteratively using the learned linear function. The effectiveness of the proposed method is proved on a mobile phone captured palm image dataset.