基于优化卷积神经网络结构的交通标志识别

发布时间：2018-06-12 20:58

本文选题：卷积神经网络 + 批量归一化　；参考：《计算机应用》2017年02期

【摘要】：现有算法对交通标志进行识别时,存在训练时间短但识别率低,或识别率高但训练时间长的问题。为此,综合批量归一化(BN)方法、逐层贪婪预训练(GLP)方法,以及把分类器换成支持向量机(SVM)这三种方法对卷积神经网络(CNN)结构进行优化,提出基于优化CNN结构的交通标志识别算法。其中:BN方法可以用来改变中间层的数据分布情况,把卷积层输出数据归一化为均值为0、方差为1,从而提高训练收敛速度,减少训练时间;GLP方法则是先训练第一层卷积网络,训练完把参数保留,继续训练第二层,保留参数,直到把所有卷积层训练完毕,这样可以有效提高卷积网络识别率;SVM分类器只专注于那些分类错误的样本,对已经分类正确的样本不再处理,从而提高了训练速度。使用德国交通标志识别数据库进行训练和识别,新算法的训练时间相对于传统CNN训练时间减少了20.67%,其识别率达到了98.24%。所提算法通过对传统CNN结构进行优化,极大地缩短了训练时间,并具有较高的识别率。
[Abstract]:The existing algorithms for traffic sign recognition have the problem of short training time but low recognition rate or high recognition rate but long training time. Therefore, the convolution neural network (CNN) structure is optimized by synthesizing batch normalized BN method, layer by layer greedy pre-training (GLP) method and replacing classifier with support vector machine (SVM). A traffic sign recognition algorithm based on optimized CNN structure is proposed. In this method, the ratio BN method can be used to change the data distribution in the middle layer, and the output data of convolution layer can be normalized to mean value of 0 and variance of 1, so as to improve the speed of training convergence, and to reduce the training time, the GLP method is to train the first layer convolution network first. After training, keep the parameters, continue to train the second layer, keep the parameters until all the convolution layers have been trained, which can effectively improve the recognition rate of convolution network and the SVM classifier only focuses on the samples that are wrong in classification. Samples that have been correctly classified are no longer processed, thus improving the training speed. Using the German traffic sign recognition database to train and recognize, the training time of the new algorithm is 20.67 less than that of the traditional CNN, and the recognition rate reaches 98.244.The training time of the new algorithm is higher than that of the traditional CNN. By optimizing the traditional CNN structure, the proposed algorithm greatly reduces the training time and has a high recognition rate.
【作者单位】：哈尔滨理工大学自动化学院;中国科学院自动化研究所;
【基金】：国家自然科学基金资助项目(61573357,61503382,61403370,61273267)~~
【分类号】：TP391.41;TP183

【相似文献】