改进DeepLabv3+模型的混凝土坝表观裂缝特征提取方法

Concrete dam apparent crack characteristic extraction based on improved DeepLabv3+model

为了解决混凝土坝环境复杂造成现有算法裂缝检测难度大、效果差的问题,提出了一种改进DeepLabv3+模型的混凝土坝裂缝特征提取方法.该方法以轻量型网络替换原始骨干网络提取图像特征,降低模型复杂度;扩充空洞空间金字塔池化模块,提升编码器感受野;采用多尺度特征融合策略,提高边缘信息利用率;优化模型损失函数,克服像素不均衡的困难.采用自制混凝土坝表观裂缝图像数据集对提出方法的有效性和优越性进行了验证与评估,结果表明:构建的改进网络能准确地实现复杂背景下混凝土坝表观裂缝特征的提取,分割裂缝图像的交并比与像素精度分别为72.85%与85.36%,裂缝分割效果也明显优于其他方法,可为长期混凝土坝面裂缝监测提供有效的技术手段.

To address the problems of difficulty and poor effect of existing crack detection algorithm caused by complex environment of concrete dams,a characteristic extraction method for the cracks of concrete dams based on the improved DeepLabv3+model was proposed.This method replaced the original backbone network with lightweight network to extract image features,which reduced the complexity of the model.The atrous spatial pyramid pooling module was expanded to widen the encoder receptive field.A multi-scale feature fusion strategy was adopted to improve the utilization of edge information.Moreover,the loss function of the model was optimized to overcome the pixel imbalance.The effectiveness and superiority of the proposed method were verified and evaluated using the self-made concrete dam apparent crack image dataset.The experimental results demonstrate that the proposed network can accurately retrieval characteristics of concrete dam apparent cracks under complex background.The intersection over union and pixel accuracy of segmented crack images are 72.85%and 85.36%,respectively.Compared with other classical image segmentation models,the proposed network has a significantly more prominent crack detection effect.It can provide an effective technical support for long-term concrete dam apparent crack monitoring.