|
In this study, an improved AlexNet and transfer learning architecture was used to construct a signature recognition model based on a small number of samples to verify offline signatures. To enhance the features of signatures, drawing boards were used to extract the pen pressure and brush stroke of these signatures. The signatures of eight individuals were used to build the samples and test the generalizability of the model for signatures of different individuals. Visualization technology was first used to adjust the architecture, activation function, and normalization of the model, improving the accuracy from 77.50 to 96.87%. Subsequently, transfer learning was used to solve the problem of the small number of samples, and the number of channels of the model was changed to improve its feature extraction. After all samples were prepared, 560 samples of the source database and 80 samples of the target database were used for transfer learning, which further improved the accuracy to 97.28%. To verify whether the model can recognize newly collected signatures, several original signatures, simple forged signatures, and skilled forged signatures were tested, revealing recognition rates of 91.25%, 95.63%, and 85.63%, respectively. In addition, the overall precision, recall, and F1 score of the model were 90.68%, 91.25%, and 90.97%, respectively, confirming an improvement in the feature extraction capability by extracting brush stroke and that transfer learning can be used with a small number of samples to create a model with excellent recognition rates and generalizability.
|
