| 李国强,付乐,查琳琳,王天雷.基于多特征融合注意力机制的交通标志检测[J].高技术通讯(中文),2022,32(11):1178~1187 |
| 基于多特征融合注意力机制的交通标志检测 |
| Traffic sign detection based on multi-feature fusion attention mechanism |
| |
| DOI：10.3772/j.issn.1002-0470.2022.11.008 |
| 中文关键词: 交通标志检测； YOLO v4； 注意力机制； 残差学习； 特征融合 |
| 英文关键词: traffic sign detection, YOLO v4, attention mechanism, residual learning, feature fusion |
| 基金项目: |
| 作者 | 单位 | | 李国强 | （燕山大学电气工程学院秦皇岛 071000） | | 付乐 | （燕山大学电气工程学院秦皇岛 071000） | | 查琳琳 | （燕山大学电气工程学院秦皇岛 071000） | | 王天雷 | （燕山大学电气工程学院秦皇岛 071000） |
|
| 摘要点击次数: 760 |
| 全文下载次数: 800 |
| 中文摘要: |
| 针对复杂环境下交通标志目标检测尺寸较小、特征不明显等问题，在YOLO v4算法的基础上，提出了基于多特征融合注意力机制的方法，形成reSE-BYOLO v4算法。reSE-BYOLO v4算法首先利用高分辨率下的特征信息改进检测层部分，提高小目标的检测能力；在特征融合部分利用同一尺度特征信息构建横向连接，使模型在不增加成本情况下融合更多的特征；利用reSE模块对有效特征进行增强，该模块对SE模块进行优化，利用残差学习的思想对通道维度重新标定。在TT100K数据集上的测试结果表明，reSE-BYOLO v4算法相对YOLO v4算法平均精度(mAP)提高了6.57%，平均精确度提高了4.78%，平均召回率提高了5.26%，reSE-BYOLO v4算法对交通标志识别能力更强。 |
| 英文摘要: |
| Aiming at the problems of small size and inconspicuous feature of traffic sign target detection samples in a complex environment, based on the YOLO v4 algorithm, a method based on multi-feature fusion attention mechanism is proposed to form reSE-BYOLO v4 algorithm. The reSE-BYOLO v4 algorithm firstly improves the detection layer by using the feature information in high resolution to improve the detection ability of small targets. In the feature fusion part, the same scale feature information is used to construct the horizontal connection, so that the model can fuse more features without increasing the cost. The effective features are enhanced by reSE module, which optimizes SE module and recalibrates channel dimension by residual learning. The test results on TT100K dataset show that compared with YOLO v4 algorithm, reSE-BYOLO v4 algorithm has improved mean average precision (mAP) by 6.57%, average accuracy by 4.78%, and average recall by 5.26%. It can be seen that reSE-BYOLO v4 has a stronger ability to recognize traffic signs. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
