Combining multi‐wavelet and CNN for palmprint recognition against noise and misalignment

Abstract

A palmprint recognition approach based on multi‐wavelet and convolutional neural network (CNN) against noise and misalignment is given. CNN method has high robustness in biometrics, but a large number of training samples are necessary. Moreover, the gathered palmprint images should be cropped to obtain their region of interest (ROI) and noise pollution and misalignment are not well solved. Therefore, the original training database is augmented to reduce the effects of noise and misalignment. First, an original training palmprint image is split into five new images, and every new image is decomposed once by multi‐wavelet. Three lower frequency bands in the low‐frequency multi‐wavelet component corresponding to pre‐filters are extracted as three samples. Furthermore, the split image is downsampled as a new sample. Second, the CNN model is constructed based on the augmented database by experiments. Third, the softmax method is used to classify the test samples. At last, the final result is obtained from 20 results by using the voting method. The experimental results based on PolyU, CASIA, and IIT Delhi Touchless Palmprint Database palmprint databases show that the proposed method can effectively recognise palmprint with high robustness while there is noise and misalignment, and has a generalisation to other palmprint databases.