An unsupervised lightweight network for multispectral palmprint recognition

Abstract In this article, we propose an unsupervised convolutional deep learning network with a single layer for multispectral palmprint recognition. We refer to this method as GradNet because it depends on the magnitude and direction of the response from data-driven filters. GradNet generates a filter bank from training images using DCT. Then, there a twofold layer, which comprises two operations namely convolution using learned filters and computation of computation of gradient images (Magnitude and Direction). The binary hashing process can effectively and efficiently eliminate the over-fitting by combining different filter responses into a single feature map. The histograms of gradient magnitude, direction images has been constructed according to the feature map. The histograms of gradient magnitude, direction and single feature map are then normalized, using power-L2 rule, to cope with illumination disparity and combined. The holistic feature extraction method has been employed to attain salient characteristics. Finally, for the features matching the cosine Mahalanobis distance has been used for multispectral palmprint recognition. The proposed system has been evaluated on publicly available multispectral palmprint database of the Hong Kong Polytechnic University. Experimental analyses show that the proposed method demonstrate that our method is capable of competing with many existing state-of-the-art multispectral palmprint recognition techniques as well as outperforming many others.