应用双通道卷积神经网络的交通标识识别方法

Traffic sign recognition method based on dual channel CNN

针对交通标识识别问题,传统的LeNet-5网络结构对于交通标识识别准确率低,识别速度慢,并且忽略天气等自然因素的影响。通过卷积神经网络技术,提出一种基于LeNet-5改进的双通道、多尺度的网络结构模型。在双通道结构中每个通道包含两个分支结构,且每个通道的卷积个数和图像尺度不同,通过不同尺度图像的特征提取,使图像特征变得更为丰富。其次,改进后的网络结构大大增加卷积核的个数。最后,通过更改Sigmoid激活函数为ReLu激活函数,更改随机梯度下降算法为Adam算法,并添加Dropout层来防止过拟合,从而提高交通标识识别率。改进网络的识别率为98.6%,上下浮动0.5%,相对与传统的LeNet-5网络结构,识别率提高15%以上,验证得出改进的网络结构具有一定的鲁棒性。

Aiming at the problem of traffic sign recognition,the traditional LeNet-5 network structure has low accuracy and slow recognition,and ignores the influence of natural factors such as weather.Through convolution neural network technology,an improved two channel and multi-scale network structure model based on LeNet-5 is proposed.In the dual channel structure,each channel contains two branch structures,and the number of convolutions of each channel is different from the image scale.Through the feature extraction of images with different scales,the image features become richer.Secondly,the improved network structure greatly increases the number of convolution cores.Finally,change the Sigmoid activation function to Relu activation function,change the random gradient descent algorithm to Adam algorithm,and add dropout layer to prevent over fitting,so as to improve the traffic sign recognition rate.The recognition rate of the improved network is 98.6%,floating up and down by 0.5%.Compared with the traditional LeNet-5 network structure,the recognition rate is increased by more than 15%.It is verified that the improved network structure has certain robustness.