基于卷积神经网络的路面裂缝识别方法研究

Research on pavement crack identification method based on convolution neural network

文章以路面裂缝作为研究对象,将语义分割算法作为主力工具对其识别,对道路裂缝图像的预处理、数据集的扩充、改进经典UNet网络模型展开了分析。为了测试UNet网络的最佳深度,需要进行广泛的网络架构搜索或低效的集成测试,提出了使用MobileNetV3网络来替换UNet的编码器部分,用于特征提取,在提高效果的同时提高速度。改进的UNet-MobileNetV3网络模型相比经典的UNet模型参数量有所减少,算法运行时间缩短,结构更优化,可顺利完成裂缝的识别工作,为道路的维护和保养提供了新思路。

This paper takes pavement cracks as the research object,uses semantic segmentation algorithm as the main tool to identify them,and analyzes the preprocessing of pavement cracks image,the expansion of data set,and the improvement of classical UNet network model.To determine the optimal depth of UNet network,extensive network architecture search or inefficient integration test is needed.MobileNetV3 network is proposed to replace the encoder part of UNet for feature extraction,which improves the effect and speed.Compared with the classical UNET model,the improved UNET-MobileNetV3 network model has fewer parameters,shorter running time,and a more optimized structure.The identification of cracks has been successfully completed,which provides a new idea for road maintenance.