基于Swin Transformer的沥青路面病害分类检测研究

针对传统卷积神经网络模型在沥青路面病害检测中识别长距离裂缝结构能力不足以及面临的精度局限问题，引入Swin Transformer模型进行沥青路面病害分类研究。首先对于路面检测车采集到的沥青路面扫描图像对比度低的问题，使用直方图均衡技术处理图像，增加图像可视化效果。其次，选取三种经典卷积神经网络模型作为对比模型，并在训练过程中采用更换损失函数，调整预训练模型等手段解决过拟合问题。并选用准确率、查全率、F1- score作为评价指标。在最终实验结果中Swin Transformer识别准确率达到了80.6%，F1-score达到了0.776，不仅在整体分类准确率上超越了传统CNN模型，并且对具有长距离特征结构的病害方面具有更高的识别准确率，同时具有良好的可靠性。

Research on Classification and Detection of Asphalt Pavement Diseases Based on Swin Transformer

Aiming at the insufficient ability of the traditional convolutional neural network model to identify long-distance crack structure and the accuracy limitation in the detection of asphalt pavement disease, the Swin Transformer model was introduced to study the classification of asphalt pavement disease. First of all, for the problem of low contrast of the asphalt pavement scanning image collected by the road inspection vehicle, the histogram equalization technology is used to process the image to increase the image visualization effect. Secondly, three classic convolutional neural network models are selected as comparison models, and methods such as replacing the loss function and adjusting the pre-training model are used to solve the over-fitting problem during the training process. And select accuracy rate, recall rate, F1-score as the evaluation index. In the final experimental results, the recognition accuracy of Swin Transformer reached 80.6%, and the F1-score reached 0.776, which not only surpassed the traditional CNN model in overall classification accuracy, but also had a higher recognition of diseases with long-distance characteristic structures accuracy and good reliability.