| 摘要: |
| 针对现有关于车载限速牌识别算法所存在的检测速度慢、准确率低、无法应用于嵌入式系统等问题，提出了一种基于网络的实时限速牌识别算法。该算法基于SSD_MobileNet_v1网络框架进行改进，对原来的网络进行架构裁剪以去除冗余结构；同时引入了特征金字塔网络结构，并使用focal loss作为网络训练的分类损失。实验表明，提出的识别算法准确率可达88.11%，虽然略低于目前主流目标检测算法的检测精度，但是网络的每秒帧率(Frame per Second,FPS)可以达到35.13，拥有较快的检测速度，而权重文件只有24 MB 。因此，与其他算法相比，该算法不仅适合小型的嵌入式人工智能（Artifical Intelligence,AI）设备，而且更贴近真实车载场景下的识别。 |
| 关键词: 限速标志识别 神经网络 特征金字塔网络 嵌入式人工智能设备 |
| DOI： |
|
| 基金项目: |
|
| A real-time speed limit sign recognition algorithm based on network |
| DAI Shaosheng,WU Yunduo,XIONG Kun,XIAO Jiawei |
| (School of Communication and Information Engineering，Chongqing University of Posts and Telecommunications，Chongqing 400065,China) |
| Abstract: |
| For the problems of slow detection speed,low accuracy,and inapplicability to embedded systems in the existing vehicle speed limit sign recognition algorithms,a network-based real-time speed limit sign recognition algorithm is proposed.The algorithm is improved based on the SSD_MobileNet_v1 network framework,and the original network is tailored to remove redundant structures.At the same time,the feature pyramid network structure is introduced,and focal loss is used as the classification loss of the training network.Experiments show that the accuracy of the proposed recognition algorithm can reach 88.11%.Although it is slightly lower than the detection accuracy of current mainstream target detection algorithms,the frame per second(FPS) of the network can reach 35.13,with a faster detection speed,and the weight file is only 24 MB.Therefore,compared with other algorithms,the proposed algorithm is not only suitable for small-scale embedded artifical intelligence(AI) devices,but also closer to the recognition in real car scenes. |
| Key words: speed limit sign recognition neural network feature pyramid network embedded AI device |
