Experimental research on knuckle pattern recognition algorithm based on transfer learning

Aiming at the cumbersome feature extraction of traditional machine learning algorithms and the common problems of single feature and low recognition accuracy, a convolutional neural network model that can extract image features by itself is used to conduct experimental research on the task of knuckle pattern recognition and classification. After pre-processing the knuckle pattern image, design a pre-training network model of the Vgg-16 architecture to perform migration learning on the knuckle pattern data set. In order to improve the portability and overfitting of the pre-training network, the pre-training network Part of the layer thawing and the classifier are jointly trained, the pre-training network model is fine-tuned and optimized, and the parameter abandonment method is embedded. The analysis of the experimental results shows that the pre-training network model is more suitable for the recognition and classification tasks of this experiment after fine-tuning and optimization. The recognition accuracy of the network model on the test set is further increased, which enhances the generalization ability of the pre-training network in this experiment The analysis of the experimental results shows that the pre-training network model is more suitable for the recognition and classification tasks of this experiment after fine-tuning and optimization. The recognition accuracy of the network model on the test set is further increased, which enhances the generalization ability of the pre-training network in this experiment.