为有效解决混凝土裂缝图像识别出现的断裂和内部空腔问题,提高裂缝整体区域定位识别精度,本文提出了基于卷积长短期记忆网络(Long Short Term Memory networks,LSTM)的裂缝识别方法.通过滑动窗口切分裂缝图像使邻近裂缝间呈现时空延续性;基于编解码的图像分割思想,构建基于VGG(Visual Geometry Group)骨干网络的特征提取编码器,结合卷积LSTM模块学习裂缝的上下文关联特征,通过解码器和分类模型实现裂缝分割,建立编解码特征独立的EDConvLSTM(Encoder-Decoder Convolutional Long Short Term Memory networks)裂缝分割模型,并进一步构建编码器与解码器特征融合的FEDConvLSTM(Fused Encoder-Decoder Convolutional Long Short Term Memory networks)模型,将高层特征与底层特征相结合,在保证裂缝完整性的同时充分挖掘裂缝的边缘信息,实现混凝土裂缝的精准分割.利用Github平台Yhlleo提供的开放基准数据集DeepCrack对模型进行训练并测试,结果表明,基于VGG16骨干网络的EDConvLSTM模型在测试集上的召回率可达 86%,优化后的基于VGG19的FEDConvLSTM模型分割交并比相较于Segnet、Unet、AttentionUnet模型分别提升了6%、4%、1%.结合卷积LSTM网络的编解码分割算法能够解决裂缝识别完整性问题,并提升裂缝的识别精度.
Concrete Crack Image Recognition Technology Based on Convolutional LSTM
To effectively solve the problems of fractures and internal cavities in concrete crack image recognition,and improve the accuracy of overall crack location recognition,this paper proposed a crack recognition method based on Convolutional Long Short Term Memory Networks(LSTM).Splitting crack images through sliding windows could present spatiotemporal continuity between adjacent cracks.Based on the image segmentation concept of encoding and decoding,a feature extraction encoder based on the VGG(Visual Geometry Group)backbone network was constructed.By combining the convolutional LSTM module to learn the contextual features of cracks,crack segmentation was achieved through a decoder and classification model,and an EDConvLSTM(Encoder-Decoder Convolutional Long Short Term Memory networks)crack segmentation model with independent encoding and decoding features was established,and further a FEDConvLSTM(Fused Encoder-Decoder Convolutional Long Short Term Memory networks)model that integrates encoder and decoder features was constructed.Combining high-level features with low-level features,while ensuring the integrity of cracks,the edge information of cracks was fully extracted,and precise segmentation of concrete cracks were achieved.Using the open benchmark dataset DeepCrack provided by the Github platform Yhlleo,the model was trained and tested.The results showed that the EDConvLSTM model based on the VGG16 backbone network achieved a recall rate of 86%on the test set.The optimized FEDConvLSTM model based on VGG19 achieved a segmentation intersection improvement of 6%,4%,and 1%compared to the Segnet,Unet,and AttentionUnet models,respectively.The encoder decoder segmentation algorithm combined with convolutional LSTM network can solve the problem of crack recognition integrity and improve the accuracy of crack recognition.
万方数据
维普
