Lightweight Attention Mechanism Via Knowledge Distillation For Vein Recognition

Internal biometric identification systems based on convolutional neuron network (CNN) is precise but there are disadvantages of demanding enormous data and endless calculating. Therefore, it is difficult for CNN to be used widely. To solve the problem, we present a light CNN with a channel shuffle that automatically balances channel attention and spatial attention. Then, the teacher-student training strategy is used to solve the insufficient generalization ability of light CNN and enhance the accuracy as the model focused on the program’s information. The experiment results showed that the proposed light CNN model increased the accuracy by 14.35% without additional parameters based on the public dataset FVUSM.