基于YOLOv5-EA的交通标志识别

Traffic Sign Recognition Based On YOLOv5-EA

针对目前的交通标志识别模型检测速度慢、精通过的问题,提出了基于YOLOv5-EA的交通标志识别算法。首先选择YOLOv5作为基础模型,根据交通标志尺寸小的特点,引入了有效通道注意力机制(Efficient Channel Attention),不仅避免降维和跨通道交互保持性能,还显著降低了模型的复杂度,提高了特征提取的能力;其次通过增加小尺度检测层,提高模型小目标检测的能力;最后在骨干网络中使用BSConv代替了正则卷积,减少了模型的参数。实验结果表明,在公开的TT100K数据集的基础上进行调整后,对改进前后的模型进行训练对比,改进后YOLOv5-EA模型的mAP为87%,较原始的YOLOv5模型提升了3.7%,训练中的损失降低了34%,能够更快速、准确的检测到交通标志。

To solve the problem of slow detection speed and precise passage of current traffic sign recognition models,a YOLOv5-EA based traffic sign recognition algorithm is proposed.Firstly,YOLOv5 is selected as the base model and Efficient Channel Attention(ECA)mechanism is introduced according to the small size of traffic signs,which not only avoids the degradation and cross-channel interaction to maintain the performance,but also significantly reduces the complexity of the model and improves the capability of feature extraction;secondly,by adding a small-scale detection layer,the model improves the capability of small target detection capability by adding small-scale detection layers;finally,BSConv is used in the backbone network instead of regular convolution,which reduces the parameters of the model.The experimental results show that after adjusting on the basis of the publicly available TT100K dataset and comparing the improved model before and after training,the mAP of the improved YOLOv5-EA model is 87%,which is a 3.7%improvement over the original YOLOv5 model,and the loss in training is reduced by 34%,which can detect traffic signs more quickly and accurately.