基于改进SSD的自然场景小交通标志检测
SMALL TRAFFIC SIGN DETECTION IN NATURAL SCENE BASED ON IMPROVED SSD
郭烊君 1雷景生2
作者信息
- 1. 上海电力大学计算机科学与技术学院 上海 201300
- 2. 浙江科技学院信息与电子工程学院 浙江杭州 310023
- 折叠
摘要
为提高在复杂的自然交通场景下对小交通标志检测的准确度,改进了 SSD模型.在SSD多个检测层加入并行多尺度特征融合,通过结合深、浅特征层的检测优势,改善了 SSD模型在小目标检测方面的不足;在SSD模型的多个检测头分别加入注意力机制模块,增强对小交通标志的特征提取效果;加入focal loss损失函数减小背景对整体损失的贡献,防止背景过拟合.实验结果表明,在复杂自然场景下,改进的方法相比原始模型对小交通标志检测的mAP提升了 4.9百分点.
Abstract
In order to improve the accuracy of the of small traffic signs detection in complex natural traffic scenes,an improved SSD algorithm is proposed.Parallel multi-scale feature fusion was added to multiple detection layers of SSD.The combination of shallow and deep features compensated the disadvantages of the SSD model in its detections of the small targets.The attention mechanism was added in multiple detection heads of SSD model to enhance the feature extraction effect of small traffic signs.The focal loss function was applied to reduce the effect of the background to the overall loss and avoid overfitting to the background.The experimental results show that the mPA of detecting small traffic signs with the improved SSD model in the complex natural scenes is improved by 4.9 percentage points compared with the original model.
关键词
SSD模型/小交通标志检测/多尺度特征融合/注意力机制Key words
SSD model/Small traffic sign detection/Multi-scale feature fusion/Attention mechanism引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据