针对交通标志只在图像中占极小的区域且难以准确识别的问题,提出一种基于注意力模型的无锚框交通标志识别算法,利用密集连接网络DenseNet-121作为骨干网络并对特征进行提取。为了解决小型交通标志准确率低的问题,在骨干网络中加入注意力模型,可以对特征图进行空间和通道上的自适应调整,通过加强或抑制特征图中元素的权重可以提升对小型交通标志的识别性能。为了减小编码路径与解码路径间的语义鸿沟,引入残差网络的连接方式并提出一种语义连接路径。为了解决锚框中正负样本不均衡的问题,采用无锚框的检测方式可以定位交通标志的中心点、回归边界框的位置与尺寸信息。对所提算法在TT100K数据集上进行验证,实验结果证明所提算法具有优越性。

Aiming at the problem that traffic signs only occupy a very small area in the image and are difficult to accurately identify, an anchorless frame traffic sign recognition algorithm based on the attention model is proposed. The densely connected network DenseNet-121 is used as the backbone network and features are extracted. In order to solve the problem of low accuracy of small traffic signs, an attention model is added to the backbone network to make adaptive adjustments to the space and channel of the feature map. The recognition performance of small traffic signs can be improved by strengthening or suppressing the weight of elements in the feature map. In order to reduce the semantic gap between the encoding path and the decoding path, the residual network connection method is introduced and a semantic connection path is proposed. In order to solve the problem of the imbalance of positive and negative samples in the anchor frame, the detection method without anchor frame can locate the center point of the traffic sign to regression the position and size information of the boundary box. The proposed algorithm is verified on the TT100K dataset, and the experimental results prove the superiority of the proposed algorithm.