基于Transformer的道路病害高精度算法

High Precision Algorithm of Road Disease Based on Transformer

摘要:

为解决目前道路病害检测算法精度低、鲁棒性差的问题，提出了Bi-Swin Transformer结构。其主干特征提取网络采用四层Swin Transformer Block来捕获全局依赖关系，在特征融合阶段使用BiFPN融合网络实现特征在不同尺度之间的双向交流。在来自7个不同国家的道路病害数据库上进行训练，并应用数据增强方式。结果表明，Bi-Swin Transformer有着比Transformer模型及卷积神经网络更为优秀的性能，更适用于道路病害检测，平均准确率达到了0.539。其在小目标道路病害中的检测精度仍然为最优，高于当前先进的检测模型。通过实际检测表明，Bi-Swin Transformer模型不仅具有高精度，而且具有强鲁棒性。

Abstract:

To solve the problems of low precision and poor robustness of current road disease detection algorithm, a Bi-Swin Transformer structure is proposed. Its backbone feature extraction network employs a 4-layer Swin Transformer Block to capture the global dependencies. In the feature fusion stage, a BiFPN fusion network is used to enable the bidirectional communication of features among the different scales. A data enhancement approach is utilized by training on road disease databases from seven different countries. The results indicate that the Bi-Swin Transformer outperforms both the Transformer model and the Convolutional Neural Network, which is more suitable for the detection of road disease. Its average accuracy is 0.539. Its detection accuracy for small-target road diseases is still optimal and higher than the current advanced detection models. The Bi-Swin Transformer model has not only high precision, but also strong robustness through practical detection.