Densely Connected Convolutional Network Optimized by Genetic Algorithm for Fingerprint Liveness Detection
Jian Wen, Yujie Zhou, Hongming Liu
Abstract
Fingerprint liveness detection is an essential module for an accurate and reliable fingerprint identification system. In this paper, a Densely Connected Convolutional Network (DenseNet) is used for fingerprint liveness detection and the genetic algorithm is adopted to optimize the DenseNet structure. Firstly, all images in the experimental database are unified to the same size through ROI extraction based on thinning images, and then used as input data for subsequent classifiers. Secondly, a variable-length real array is subdivided into four gene fragments to characterize the DenseNet structure. We design specific mutation and crossover operators for the evolution of DenseNet population. The optimal structure is found from a large solution space comprising 1.4 <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">*</sub> 10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">19</sup> candidates by genetic algorithm after 30 generations of evolution. Finally, the optimal DenseNet model is compared with other state-of-the-art works in detail. The proposed model achieves 98.22% accuracy on the testing set of mixed Livdet dataset. The experimental results show that genetic algorithms can automatically find the optimal structure from the solution space and further exploit the potential of DenseNet, which can help researchers to quickly construct high-performance network structures even if they are not proficient in neural networks. By comparing the average classification error (ACE) value and variance, it can be concluded that the classification performance of the proposed model is more accurate and balanced than other state-of-the-art models.