摘要
为准确检测混凝土路面裂缝的形态与分裂程度,避免其结构进一步受损,提出了一种改进的DeepLabV3网络语义分割模型。利用Canny算法优异的检测能力对裂缝边缘进行提取,改进分割网络的上采样层进行残差多层采样;优化空洞卷积的扩张率降低感受野,平衡网络对不同尺度裂缝的敏感度;融合并行注意力模块抑制分割模型易产生的伪影效应,获取更具互补性的裂缝特征。在公开数据集上进行训练与预测,在全卷积网络(FCN)结合条件随机场(CRF)方法、Deep LabV3方法、Deep LabV3+方法与Lraspp方法中开展了对比实验。实验结果表明,本方法的MPA为98.73%,MIOU为87.53%,有效抑制噪声干扰,分割结果精确且连续。
To accurately detect the morphology and splitting degree of concrete pavement cracks and avoid further structural damage,an improved semantic segmentation model of DeepLabV3 network was proposed.Firstly,the crack edge is extracted by the excellent detection ability of Canny algorithm,and improve network segmentation on sampling for residual multilayer sampling.Secondly,the expansion rate of cavity convolution was optimized to reduce the receptive field and balance the sensitivity of the network to cracks of different scales.Finally,the convolutional block attention module is integrated to suppress the artifact effect easily generated by the segmentation model and obtain more complementary fracture features.The proposed method is trained and tested on open data sets,and compared with the full convolutional network(FCN)combined with conditional random field(CRF)algorithm,DeepLabV3 algorithm,Deep LabV3+algorithm and Lraspp algorithm.The experimental results show that the MPA and MIOU of this method are 98.73%and 87.53%respectively,which can effectively suppress noise interference and obtain accurate and continuous segmentation results.
作者
张卫国
张思瑞
ZHANG Wei-guo;ZHANG Si-rui(College of computer science&technology,Xi’an University of Science and Technology,Xi’an,Shaanxi 710054,China)
出处
《计算技术与自动化》
2023年第3期96-101,共6页
Computing Technology and Automation
基金
国家自然科学基金青年科学基金项目(61902311)。
关键词
图像处理
语义分割
裂缝检测
全卷积网络
CANNY边缘检测
image processing
semantic segmentation
crack detection
full convolutional network
Canny operator
