基于Transformer的交通标志检测模型研究

Research on Traffic Sign Detection Model Based on Transformer

【目的】为了解决在复杂环境下,对小目标特征困难以及对小目标检测效果不佳等问题,提出了一种基于Transformer的交通标志检测基干模型。【方法】通过充分利用卷积和Transformer的优势,构建了一种注意力融合的多尺度特征提取基干模型,能够使基干网络以全局上下文信息为支撑,有选择地增强有用信息的特征,并抑制不重要的特征。此外,为了在增强特征融合的同时防止网络退化,还加入了类池连接。最后,在TT100K数据集上进行实验。【结果】实验结果表明,以该模型为骨干的元体系结构取得了最高84%的mAP,与基线模型相比m AP最大提升约7%。【结论】模型在提高特征提取效果的同时,也为交通标志检测提供了一种新的思路。

【Objective】In order to solve the difficulties such as small target feature extraction,a transformerbased traffic sign detection model was proposed.【Method】Through fully utilizing the advantages of convolution and Transformer,a multi-scale feature extraction backbone model was established with attention fusion,which could enable the backbone network to selectively enhance the features of useful information and suppress the unimportant ones with the support of global context information.In addition,pooling-like connection are incorporated in order to prevent network degradation while enhancing feature fusion.Finally,experiments were conducted on the TT100K dataset.【Result】The experimental results show that the meta-architecture with this model as the backbone achieves the highest mAP of 84%,and the maximum improvement of mAP is about 7%compared with the baseline model.【Conclusion】The model provides a new idea for traffic sign detection while improving feature extraction.