In recent years, target detection framework based on deep learning has made brilliant achievements. However, real-life traffic sign detection remains a great challenge for most of the state-of-the-art object detection methods. The existing deep learning models are inadequate to effectively extract the features of small traffic signs from large images in real-world conditions. In this paper, we address the small traffic sign detection challenge by proposing a novel small traffic sign detection method based on a highly efficient end-to-end deep network model. The proposed method features fast speed and high precision as it attaches three key insights to the established You Only Look Once (YOLOv3) architecture and other correlated algorithms. Besides, network pruning is appropriately introduced to minimize network redundancy and model size while keeping a competitive detection accuracy. Furtherly, four scale prediction branches are also adopted to significantly enrich the feature maps of multi-scales prediction. In our method, we adjust the loss function to balance the contribution of error source to the total loss. The effectiveness, and robustness of the network is further proved with experiments on Tsinghua-Tencent 100 K traffic sign dataset. The experimental results indicate that our proposed method has achieved better accuracy than that of the original YOLOv3 model. Compared with the schemes in relevant literatures our proposed method not only emerges performance superiors in detection recall and accuracy, but also achieves 1.9–2.7x improvement in detection speed.

近年来，基于深度学习的目标检测框架取得了辉煌的成就。然而，对于大多数最新的物体检测方法来说，现实交通标志检测仍然是一个巨大的挑战。现有的深度学习模型不足以有效地在现实条件下从大图像中提取小交通标志的特征。在本文中，我们通过提出一种基于高效端到端深度网络模型的新颖的小交通标志检测方法来解决小交通标志检测的挑战。所提出的方法具有快速和高精度的优点，因为它对已建立的“一次只看一次”（YOLOv3）体系结构和其他相关算法具有三个关键见解。除了，适当引入了网络修剪功能，以最大程度地减少网络冗余和模型大小，同时保持具有竞争力的检测精度。此外，还采用了四个尺度预测分支来显着丰富多尺度预测的特征图。在我们的方法中，我们调整损失函数以平衡误差源对总损失的贡献。通过在清华腾讯100 K交通标志数据集上进行的实验进一步证明了该网络的有效性和鲁棒性。实验结果表明，我们提出的方法比原始YOLOv3模型具有更高的精度。与相关文献中的方案相比，我们提出的方法不仅在检测召回率和准确性方面表现出优异的性能，而且在检测速度上也提高了1.9-2.7倍。