基于改进的GoogleNet-ResNet算法的路基病害智能分类方法

Intelligent Classification Method for Road Pavement Diseases Based on Improved GoogleNet-ResNet Algorithm

摘要:

针对路基病害分类算法存在的复杂病害辨识难度大、多视图雷达图像特征利用不充分等问题,本文提出一种基于改进的GoogleNet-ResNet算法的路基病害智能分类方法。首先,引入坐标注意力和改进的Inception模块对GoogleNet网络结构进行优化。然后,利用改进的GoogleNet学习c-scan数据特征剔除非目标病害,实现病害目标的粗分类。最后,将分类成病害的b-scan数据输入基于迁移学习的ResNet50,实现病害的细分类。实验表明,改进的GoogleNet进行病害粗分类的准确率可达到98.2%,检测速度可达90.9FPS。基于迁移学习的ResNet50进行病害细分类的准确率可达90.5%,检测速度可达52.6FPS。本文算法的准确率比单独的改进的GoogleNet网络高10.1%,比单独的ResNet50网络高7.4%,有效地提高了道路路基病害的识别精度与效率。

Abstract:

In response to the challenges of complex disease identification and the underutilization of multi-view radar image features in the classification algorithm for roadbed diseases, this paper proposes an intelligent classification method for roadbed diseases based on an improved GoogleNet-ResNet algorithm. Firstly, it introduces coordinate attention and improved Inception modules to optimize the GoogleNet network structure. Then, the improved GoogleNet is utilized to learn the c-scan data features, eliminating non-target diseases and achieving a coarse classification of the disease targets. Finally, the b-scan data classified as diseases is input into a ResNet50 model based on transfer learning to achieve the fine classification of the diseases. The results show that the improved GoogleNet achieves an accuracy of 98.2% for coarse disease classification and a detection speed of 90.9 FPS. The accuracy of disease sub-classification using ResNet50 based on transfer learning reaches 90.5%, and the detection speed reaches 52.6 FPS. The proposed algorithm in this paper achieves an accuracy that is 10.1% higher than that of the improved GoogleNet network alone, and 7.4% higher than that of the ResNet50 network alone. This algorithm effectively improves the recognition accuracy and efficiency of roadbed disease detection while reducing the probability of misjudgments.