融合注意力机制和新型卷积神经网络的市政道路病害识别

Municipal Road Disease Recognition Based on Attention Mechanism and Novel Convolutional Neural Network

环境和荷载协同作用导致的路面病害对道路使用性能和安全性能的影响日益突出。现有图像智能识别算法难以实现处理速度和计算量的平衡。针对道路病害快速准确实时识别的需求,对石家庄损伤较为严重的路面进行实地拍照,结合已有图片,采用数据增强技术构建了市政道路病害数据集,并且提出了一种基于MobileNetV3网络的轻量化道路病害识别网络模型GEM-MobileNetV3。该模型首先使用Ghost模块代替MobileNetV3网络基本单元中的1×1卷积;然后结合改进后的高效通道注意力机制(efficient channel attention,ECA)模块提取病害目标的重要特征;最后将网络浅层的ReLU激活函数替换为泛化能力更强的Mish激活函数,提高模型的整体性能。通过消融实验与对比实验,验证了新模型的有效性。实验结果表明,新模型准确率达到96.33%,其参数量与计算量较MobileNetV3模型分别降低了37.9%和36%。提出的新模型在保持较高识别准确率的同时有效降低了模型复杂度,为在低成本计算平台上实现高准确率实时识别提供了新途径。

Pavement diseases caused by the synergistic effect of environment and load have an increasingly prominent impact on in-service performance and durability performance.It is difficult for the existing intelligent image recognition algorithms to achieve a balance between processing speed and computational complexity.Aiming at the requirements of fast,accurate,and real-time identification of road diseases,the road surface with serious damage in Shijiazhuang was taken on the spot,combined with the existing pictures,the data augmentation technology was used to construct the municipal road disease dataset,and a lightweight road disease identification network model,GEM-MobileNetV3,was proposed based on the MobileNetV3 network.Firstly,the Ghost module was used to replace the 1×1 convolution in the basic unit of the MobileNetV3 network.Then,combined with the improved efficient channel attention module(ECA),essential features of disease targets were extracted.Finally,the ReLU activation function in the shallow layer of the network was replaced by the Mish activation function with a stronger generalization ability to improve the overall performance of the model.The effectiveness of the new model was verified by ablation experiments and comparative experiments.Experimental results show that the accuracy of the new model reaches 96.33%,and the number of parameters and computations are reduced by 37.9%and 36%,respectively,compared with the MobileNetV3 model.The proposed new model can effectively reduce computational complexity while maintaining high recognition accuracy,which provides a new way to achieve high-accuracy real-time recognition on low-cost computing platforms.