基于改进萤火虫算法的卷积神经网络在人脸识别中的研究

Study on Convolutional Neural Network in Face Recognition Based on Improved Firefly Algorithm

针对卷积神经网络采用人工方法选择超参数影响模型识别效果的问题,本文提出采用具有泛化能力强和参数少等优点的萤火虫算法对超参数进行优化的思路。首先,针对萤火虫算法容易陷入局部最优和容易收敛的缺点,提出基于混沌思想的种群初始化和自适应步长的优化策略,其次,选择6个卷积神经网络的超参数使用萤火虫算法进行优化;最后,在仿真实验中,将改进后的萤火虫算法进行性能测试,并与其他2种卷积神经网络算法在人脸数据库中进行对比,在识别率和完成时间上都获得较好的效果。

To address the problem that the hyperparameters of convolutional neural networks are selected by manual methods to affect the model recognition effect,the idea of using the firefly algorithm with the advantages of strong generalization ability and few parameters to optimize the hyperparameters is proposed.Firstly,for the disadvantages that the firefly algorithm is easy to fall into local optimum and easy to converge,the optimization strategy of population initialization and adaptive step size based on chaos idea is proposed,secondly,six hyperparameters of the convolutional neural network are selected for optimization using the firefly algorithm,finally,in the simulation experiment,the improved firefly algorithm is tested for performance and compared with other two convolutional neural network algorithms in the face database for comparison,and better results are obtained in terms of recognition rate and completion time.