基于集中式特征金字塔的交通标志识别

Traffic Sign Recognition Based on Centralized Feature Pyramid

针对目前交通标志识别技术中存在的畸变目标、小目标检测难等问题,提出一种基于集中式特征金字塔的交通标志识别算法。首先,使用集中式特征金字塔改进原始的特征融合网络,用轻量级多层感知机(MLP)来捕获全局远程依赖,通过可学习视觉中心机制(LVC)来捕获输入图像的局部角区域,提高了对畸变目标以及小目标的检测精度;其次,使用递归门控卷积提取浅层特征图的高阶空间交互信息,改善对小目标的检测效果;最后,使用SIoU回归损失函数,引入角度损失,重新定义惩罚指标,减少总损失的自由度,防止预测框在训练时四处游荡,加快收敛速度,使定位更加精确。在TT100K数据集上平均检测精度为93.4%,和传统的YOLOv5n相比精度提升了3.5个百分点,帧处理速度达到94.34fps。

A traffic sign recognition algorithm based on feature concentration pyramid was proposed to solve the problems of distorted target and small target detection.Firstly,the original feature fusion network is improved by using the feature concentration pyramid,and the global remote dependence is captured by a lightweight multi-layer perceptron(MLP).The local corner region of the input image is captured by the parallel learning vision center mechanism(LVC),which improves the detection accuracy of distorted targets and small targets.Secondly,recursive gated convolution is utilized to extract high-order spatial interaction information of shallow feature maps,which is beneficial to improve the detection effect of small targets.Finally,the SloU regression loss function is used,with angle loss introduced to redefine the penalty index.It not only reduces the degree of freedom of total loss,preventing the prediction box from wandering around during training,but also speeds up the convergence rate to make the positioning more accurate.Numerous experiments on the TT100K data set demonstrate that the average detection accuracy is 93.4%,which is 3.5% higher than that of the traditional YOLOv5n,and the frame processing speed reaches up to 94.34fps.