摘要
针对目前利用深度学习和计算机视觉技术检测混凝土裂缝需要较好光照条件的现状,实现了在低光照条件下的混凝土裂缝检测。首先,为了加快网络训练速度和检测速度,采用Segnet-MobileNet轻量级深度神经网络作为主要架构。其次,通过引入感受野模块和图像增强模块,利用图像增强模块对裂缝样本进行光照增强预处理,并将感受野模块连接到MobileNet模型后端,利用感受野模块中的残差结构、多分支结构和膨胀卷积融合多尺度裂缝特征图谱信息,扩大网络感受野,进一步改善裂缝检测效果。最后,通过取样验证、损失评价和精度评价的方法评价模型性能,并且通过进行多次试验减少结果偶然性。试验结果表明:模型具有较高的检测精确度,平均检测精确度为97.6%,解决了低光照条件下混凝土裂缝的检测问题。
In view of the current situation that deep learning and computer vision technology are used to detect concrete cracks under better light conditions, the detection of concrete cracks under low light conditions is realized. Firstly, in order to accelerate the network training speed and detection speed, Segnet-MobileNet lightweight deep neural network is adopted as the main architecture. Secondly, the receptive field block and image enhance block are used in the network.The image enhance block is used to pre-process the crack samples with light enhancement. The receptive field block is connected to the back-end of MobileNet model. The receptive field block uses residual structure, multiple branching structure and dilate convolution to fuse crack multi-scale feature maps information, and expands the receptive field of the network, further improve the effect of crack detection. Finally, the performance of the model was evaluated by sampling validation, loss evaluation and accuracy evaluation, and reduce fortuity by conducting multiple trials. The experimental results show that the model has a high detection accuracy, with an average detection accuracy of 97.6%, which solves the detection problem of concrete cracks in low light condition.
作者
李想
熊进刚
LI Xiang;XIONG Jingang(School of Civil Engineering and Architecture,Nanchang University,Nanchang 330031,China)
出处
《建筑结构》
CSCD
北大核心
2021年第S02期1046-1050,共5页
Building Structure
基金
国家自然科学基金项目(51768044、52068051)
关键词
深度学习
裂缝检测
感受野模块
混凝土
低光照条件
deep learning
crack detection
receptive field block
concrete
low light condition
