改进的深度卷积网络在交通标志识别中的应用

Application of improved depth convolution network in traffic sign recognition

[目的]交通标志的采集大多是在自然环境下进行的,因为背景干扰、视角倾斜、局部遮挡等不利条件下导致采集的图像质量不高;传统的卷积神经网络(convolution neural network,CNN)只有一条通道提取特征,在面对质量不高的图像时,会出现特征提取不充分的情况导致分类准确率不高。[方法]提出了一种改进的网络模型,该模型集合了多尺度输入、并行交叉以及恒等映射的特点。[结果]能够保证特征提取的充分性与多样性并使网络性能不会随深度加深而退化。[结论]在实验阶段用该网络对德国交通标志数据库(GTSRB)进行测试,取得了97.6%的准确率。

Most of the traffic signs are collected in the natural environment,the image quality is not high because of the adverse conditions such as background interference,angle tilt and local shelter.Because the traditional convolution neural network（CNN） has only one channel to extract features,when dealing with images of low quality,it may appear feature extraction is not sufficient,which leads to low classification accuracy.Due to the mentioned reasons,we present an improved network model,which contains the characteristics of multi-scale input,parallel crossover and identity mapping.It can ensure the sufficiency and diversity of feature extraction,and the network performance will not be degraded with depth.In the test stage,the German traffic sign database （GTSRB） was tested by this network,and an accuracy of 97.6％ is achieved.