基于优化卷积神经网络的人耳识别

Human ear recognition based on optimized convolutional neural network

为了提高人耳图像的识别率,本文提出了一种基于优化卷积神经网络的人耳识别算法。首先针对人耳识别问题设计一个基于卷积神经网络的深层网络结构,选取改进的激活函数PReLU,在最后的全连接层引入Dropout技术,防止网络过拟合,选择不依赖学习率也可以达到良好效果,且收敛速度最快的Adadelta对网络进行优化。实验中通过大量的不同类型的人耳图像样本对设计的网络不断地训练,以此来确定卷积神经网络中特征图数量以及学习率等参数的设置。对训练好的深度学习网络结构进行最后的人耳识别步骤测试,通过对比实验证明了该方法在一定程度的旋转以及遮挡等干扰条件下具有很强的鲁棒性,同时人耳识别率明显提高。

In order to improve the recognition rate of human ear images,an algorithm for recognizing human ears based on optimized convolutional neural network is proposed. Firstly,a deep network structure based on convolutional neural network is designed for human ear recognition problem. Secondly,the optimal activation function PreLU is selected and Dropout technology is introduced in the final fully connected layer to prevent network over-fitting. And the Adadelta,which can achieve good results without relying on the leaning rate and also has the fastest convergence rate,is chosen to optimize the network. In experiments,a large number of human ear images of different types are used as samples to continuously train the designed network,so as to determine the settings of the number of feature maps and the learning rate in the convolutional neural network.As a final step,human ear recognition experiments on the trained deep learning network structure are carried out. Results of the comparison experiments show that the method has strong robustness to certain interference conditions such as rotation and occlusion,etc. so that the recognition rate of the human ear is significantly improved.