Abstract: Aiming at the problem that the generalization ability of traditional convolutional neural network for multi-sensor fingerprint recognition is reduced and the accuracy is not high, an improved Stacking algorithm is proposed. Firstly, AlexNet is improved by introducing depth-separable convolution to reduce the number of parameters and speed up the training. The spatial pyramid pool is introduced to improve the ability of the network to obtain global information. Batch normalization is introduced to speed up network convergence and improve accuracy of the network on the test set. Global average pooling is used instead of fully connected layer to prevent overfitting. Then DenseNet and the improved AlexNet convolutional neural networks are used as the base learner of Stacking to classify fingerprints and obtain the prediction results. Finally, each model trained with the same base learner is weighted according to the prediction accuracy, and the prediction results are then classified by the meta-classifier. The improved Stacking algorithm is tested on multi-sensor fingerprint database, and the final recognition accuracy is 98.43%, which is 20.05% higher than AlexNet and 4.25% higher than DenseNet. 

Key words: fingerprint recognition, DenseNet, AlexNet, Stacking ensemble learning, convolutional neural network