Joint deep convolutional feature representation for hyperspectral palmprint recognition

• Published in: Information sciences Vol. 489; pp. 167 - 181
• Main Authors: Zhao, Shuping, Zhang, Bob, Philip Chen, C.L.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.07.2019
• Subjects: CNN stack; Hyperspectral palmprint recognition; Joint convolutional feature; Hyperspectral palmprint recognition; Joint convolutional feature; CNN stack
• ISSN: 0020-0255; 1872-6291
• DOI: 10.1016/j.ins.2019.03.027

•A new feature representation method for hyperspectral palmprint (HSPP) is proposed.•Proposed method is called joint deep convolutional feature representation (JDCFR).•JDCFR is a CNN stack where each stack element is a novel CNN for a unique HSPP band.•Extracted features in all bands are combined forming a joint convolutional feature.•JDCFR tested on large HSPP data for recognition achieved 0.01% EER and 99.62% ARR. With discriminative information from various spectrums, hyperspectral imaging analysis has recently attracted more and more considerable research attention. This increase can also be attributed to an improvement in computer hardware that has led to the development of Convolutional Neural Networks (CNNs) achieving very high performances in numerous applications. Motivated by these technologies, a Joint Deep Convolutional Feature Representation (JDCFR) methodology is proposed for hyperspectral palmprint recognition. For a hyperspectral palmprint image cube, a CNN stack is constructed to extract its features from the entire spectral bands and generate a joint convolutional feature. The CNN stack contains dozens of CNNs with different parameter settings, which can be trained locally using palmprint images in different spectrums. To obtain a complete and nonredundant set of features and to avoid losing hierarchical characteristics hidden in different bands, the Joint Deep Convolutional Feature is represented by a Collaborative Representation-based Classifier (CRC) simultaneously to perform classification. Experimental results were conducted on a hyperspectral palmprint dataset consisting of 53 spectral bands with 110,770 images. Compared with other classifiers, CNNs, traditional palmprint recognition methods, as well as applying PCA to the feature matrix, the proposed method achieved the highest performance with 0.01% – EER and 99.62% – ARR.