深度特征重构与权重分配的交通标志识别算法

Traffic Sign Recognition Based on Depth Feature Reconstruction and Weight Assignment

针对目前基于深度学习的交通标志识别方法在空间不变性以及深度有效特征利用率不足等方面存在的问题,提出了一种基于特征重构与权重分配的交通标志识别算法.首先,将交通标志输入预处理空间变换网络,获得具有空间不变性的图像;然后,压缩卷积特征,提取全局特征信息,得到不同通道特征对网络识别交通标志的贡献率,多层全连接学习深度特征重要性,引入缩放参数r,减少网络参数计算,并分配深度特征权重,完成特征重构;最后,确定阶段特征中卷积特征需要重构的位置,完成3个阶段特征的重标定.在公开的德国交通标志数据集(GTSRB)上进行实验,对43类12630张交通标志测试集,识别率可达到99. 32%.

Aiming at the problems existing in the current deep-learning traffic sign recognition method in terms of spatial invariance and insufficient utilization of deep effective features,a traffic sign recognition algorithm based on feature reconstruction and weight distribution is proposed. Firstly,the traffic signs are input into the pre-processed spatial transformation network to obtain spatially invariant images. Then,the deep features are compressed,and the global feature information is extracted to obtain the contribution rate of different channel features to the network identification traffic signs. Multi-layer full-connection learns the importance of depth features,which introduces scaling parameters r to reduce network parameters calculation,and the depth feature weight is assigned to complete the feature reconstruction. Finally,the location of the convolution feature in the phase feature needs to be reconstructed,and the re-calibration of the three phase features is completed. Experiments were conducted on the public German traffic sign data set( GTSRB),which the recognition rate of the class 43 of 12,630 traffic sign test sets can reach 99. 32%.