基于卷积神经网络的交通标志识别方法

Traffic sign recognition method based onconvolutional neural network

交通标志识别在自动驾驶过程中起着十分重要的作用。为了解决识别精度低的问题,提出一种基于卷积神经网络的识别方法,通过改进深度相互学习网络完成对交通标志的识别,使用ResNet-19网络作为特征提取部分,使用全局平均池化层作为分类器部分,使用交叉熵损失和相对熵损失作为损失函数部分,并增加超参数α与δ来衡量这两个损失在训练中的权重;同时,引入一种使用不同初始值的批量归一化层训练的技巧,以此来提高模型的收敛速度。试验研究中,改进的方法用在德国交通标志识别测试集上达到了98.90%的识别精度,比改进前精度提高了2.17%,与目前优秀的交通标志识别模型相比,本方法精度仍有一定的提高。试验结果表明在复杂的环境中,本方法可以准确地识别交通标志,这为后续相关研究提供了良好的技术支持。

Traffic sign recognition plays a very important role in the process of autonomous driving.In order to solve the problem of low recognition accuracy,a recognition method was proposed on the basis of convolutional neural network.The recognition of traffic signs was completed by improving the deep mutual learning network,consisting of the ResNet-19 network functioning as the feature extraction part,the global average pooling layer as the classifier part,cross entropy loss and relative entropy loss as the loss function part with hyperparameters added to measure the weight of these two losses in training.Meanwhile,a batch normalization was introduced by using different initial values layer training techniques to improve the convergence speed of the model.In the experimental study,the improved method was used on the German traffic sign recognition test dataset to achieve a recognition accuracy of 98.90%,2.17%higher than the accuracy prior to the improvement.Compared with the current excellent traffic sign recognition model,the accuracy of this method is still improved.The test results show that the method can accurately identify traffic signs in a complex environment,which provides sound technical support for subsequent related research.