Palmprint Classification with Deep Learning Features and SURF Keypoints

Images of the palmprint or of the entire hand can be successfully used in biometric applications. In this work, we investigate two types of methods for palmprint recognition: one uses deep learning features and the other SURF keypoints. We use several well-known pretrained Convolutional Neural Networks (CNN) to extract features, the classification process being performed with a minimum distance classifier. The keypoint approach computes the same number of SURF descriptors for all the images in the dataset. The comparison of two images is realized by an adapted keypoints matching procedure. To reduce the computing time required for the SURF based classification process for large number of descriptors, we propose an algorithm that uses both approaches. The CNN features help selecting a reduced training subset, and then, on this subset one uses the SURF procedure with a high number of computed keypoints to accurately establish the label of a test image. The experiments were performed on 5 hand/palmprint datasets: GPDS, IITD, CASIA, CASIA Multi-Spectral, PolyU.