The original paper is in English. Non-English content has been machine-translated and may contain typographical errors or mistranslations. ex. Some numerals are expressed as "XNUMX".
Copyrights notice
The original paper is in English. Non-English content has been machine-translated and may contain typographical errors or mistranslations. Copyrights notice
Sebagai kecacatan permukaan biasa struktur pelapik terowong, penyakit rekahan menjejaskan ketahanan struktur terowong dan menimbulkan bahaya tersembunyi kepada keselamatan pemanduan terowong. Faktor-faktor seperti gangguan daripada persekitaran perkhidmatan kompleks terowong dan nisbah isyarat-ke-bunyi yang rendah bagi sasaran retak itu sendiri, telah menyebabkan kaedah pengecaman retak sedia ada berdasarkan segmentasi semantik tidak dapat memenuhi keperluan kejuruteraan sebenar. Berdasarkan ini, kertas kerja ini menggunakan rangkaian Unet sebagai rangka kerja asas untuk pengecaman retak dan bercadang untuk membina model peningkatan lata lilitan berbilang kernel (MKCE) untuk mencapai pengesanan dan pengenalpastian penyakit retak yang tepat. Pertama sekali, untuk memastikan prestasi pengekstrakan ciri retak, model itu mengubah suai rangkaian pengekstrakan ciri utama dalam rangka kerja asas kepada rangkaian sisa ResNet-50. Berbanding dengan rangkaian VGG-16, pengubahsuaian ini boleh mengekstrak ciri perincian retak yang lebih kaya sambil mengurangkan parameter model. Kedua, memandangkan rangkaian Unet tidak dapat melihat dengan berkesan ciri retak berskala dalam peringkat sambungan langkau, modul peningkatan lata lilitan berbilang kernel dicadangkan dengan menggabungkan sambungan lata kumpulan lilitan berbilang kernel dan kumpulan lilitan diluaskan kadar berbilang pengembangan. . Modul ini mencapai persepsi menyeluruh tentang butiran tempatan dan kandungan global retak lapisan terowong. Di samping itu, untuk lebih melemahkan kesan gangguan kekacauan latar belakang terowong, modul pengiraan perhatian blok konvolusi selanjutnya diperkenalkan selepas modul peningkatan lata lilitan berbilang kernel, yang mengurangkan kadar penggera palsu pengecaman retak dengan berkesan. Algoritma ini diuji pada sejumlah besar set data imej retak terowong bawah tanah. Keputusan eksperimen menunjukkan bahawa, berbanding dengan algoritma pengecaman retak lain berdasarkan pembelajaran mendalam, kaedah dalam kertas ini telah mencapai keputusan terbaik dari segi ketepatan dan penunjuk persilangan atas kesatuan (IoU), yang mengesahkan kaedah dalam kertas ini mempunyai kebolehgunaan yang lebih baik. .
Baoxian WANG
Shijiazhuang Tiedao University
Zhihao DONG
Shijiazhuang Tiedao University
Yuzhao WANG
Shijiazhuang Tiedao University
Shoupeng QIN
China Railway design corporation
Zhao TAN
China Railway design corporation
Weigang ZHAO
Shijiazhuang Tiedao University
Wei-Xin REN
Shenzhen University
Junfang WANG
Shenzhen University
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Salinan
Baoxian WANG, Zhihao DONG, Yuzhao WANG, Shoupeng QIN, Zhao TAN, Weigang ZHAO, Wei-Xin REN, Junfang WANG, "Visual Inspection Method for Subway Tunnel Cracks Based on Multi-Kernel Convolution Cascade Enhancement Learning" in IEICE TRANSACTIONS on Information,
vol. E106-D, no. 10, pp. 1715-1722, October 2023, doi: 10.1587/transinf.2023EDP7073.
Abstract: As a typical surface defect of tunnel lining structures, cracking disease affects the durability of tunnel structures and poses hidden dangers to tunnel driving safety. Factors such as interference from the complex service environment of the tunnel and the low signal-to-noise ratio of the crack targets themselves, have led to existing crack recognition methods based on semantic segmentation being unable to meet actual engineering needs. Based on this, this paper uses the Unet network as the basic framework for crack identification and proposes to construct a multi-kernel convolution cascade enhancement (MKCE) model to achieve accurate detection and identification of crack diseases. First of all, to ensure the performance of crack feature extraction, the model modified the main feature extraction network in the basic framework to ResNet-50 residual network. Compared with the VGG-16 network, this modification can extract richer crack detail features while reducing model parameters. Secondly, considering that the Unet network cannot effectively perceive multi-scale crack features in the skip connection stage, a multi-kernel convolution cascade enhancement module is proposed by combining a cascaded connection of multi-kernel convolution groups and multi-expansion rate dilated convolution groups. This module achieves a comprehensive perception of local details and the global content of tunnel lining cracks. In addition, to better weaken the effect of tunnel background clutter interference, a convolutional block attention calculation module is further introduced after the multi-kernel convolution cascade enhancement module, which effectively reduces the false alarm rate of crack recognition. The algorithm is tested on a large number of subway tunnel crack image datasets. The experimental results show that, compared with other crack recognition algorithms based on deep learning, the method in this paper has achieved the best results in terms of accuracy and intersection over union (IoU) indicators, which verifies the method in this paper has better applicability.
URL: https://global.ieice.org/en_transactions/information/10.1587/transinf.2023EDP7073/_p
Salinan
@ARTICLE{e106-d_10_1715,
author={Baoxian WANG, Zhihao DONG, Yuzhao WANG, Shoupeng QIN, Zhao TAN, Weigang ZHAO, Wei-Xin REN, Junfang WANG, },
journal={IEICE TRANSACTIONS on Information},
title={Visual Inspection Method for Subway Tunnel Cracks Based on Multi-Kernel Convolution Cascade Enhancement Learning},
year={2023},
volume={E106-D},
number={10},
pages={1715-1722},
abstract={As a typical surface defect of tunnel lining structures, cracking disease affects the durability of tunnel structures and poses hidden dangers to tunnel driving safety. Factors such as interference from the complex service environment of the tunnel and the low signal-to-noise ratio of the crack targets themselves, have led to existing crack recognition methods based on semantic segmentation being unable to meet actual engineering needs. Based on this, this paper uses the Unet network as the basic framework for crack identification and proposes to construct a multi-kernel convolution cascade enhancement (MKCE) model to achieve accurate detection and identification of crack diseases. First of all, to ensure the performance of crack feature extraction, the model modified the main feature extraction network in the basic framework to ResNet-50 residual network. Compared with the VGG-16 network, this modification can extract richer crack detail features while reducing model parameters. Secondly, considering that the Unet network cannot effectively perceive multi-scale crack features in the skip connection stage, a multi-kernel convolution cascade enhancement module is proposed by combining a cascaded connection of multi-kernel convolution groups and multi-expansion rate dilated convolution groups. This module achieves a comprehensive perception of local details and the global content of tunnel lining cracks. In addition, to better weaken the effect of tunnel background clutter interference, a convolutional block attention calculation module is further introduced after the multi-kernel convolution cascade enhancement module, which effectively reduces the false alarm rate of crack recognition. The algorithm is tested on a large number of subway tunnel crack image datasets. The experimental results show that, compared with other crack recognition algorithms based on deep learning, the method in this paper has achieved the best results in terms of accuracy and intersection over union (IoU) indicators, which verifies the method in this paper has better applicability.},
keywords={},
doi={10.1587/transinf.2023EDP7073},
ISSN={1745-1361},
month={October},}
Salinan
TY - JOUR
TI - Visual Inspection Method for Subway Tunnel Cracks Based on Multi-Kernel Convolution Cascade Enhancement Learning
T2 - IEICE TRANSACTIONS on Information
SP - 1715
EP - 1722
AU - Baoxian WANG
AU - Zhihao DONG
AU - Yuzhao WANG
AU - Shoupeng QIN
AU - Zhao TAN
AU - Weigang ZHAO
AU - Wei-Xin REN
AU - Junfang WANG
PY - 2023
DO - 10.1587/transinf.2023EDP7073
JO - IEICE TRANSACTIONS on Information
SN - 1745-1361
VL - E106-D
IS - 10
JA - IEICE TRANSACTIONS on Information
Y1 - October 2023
AB - As a typical surface defect of tunnel lining structures, cracking disease affects the durability of tunnel structures and poses hidden dangers to tunnel driving safety. Factors such as interference from the complex service environment of the tunnel and the low signal-to-noise ratio of the crack targets themselves, have led to existing crack recognition methods based on semantic segmentation being unable to meet actual engineering needs. Based on this, this paper uses the Unet network as the basic framework for crack identification and proposes to construct a multi-kernel convolution cascade enhancement (MKCE) model to achieve accurate detection and identification of crack diseases. First of all, to ensure the performance of crack feature extraction, the model modified the main feature extraction network in the basic framework to ResNet-50 residual network. Compared with the VGG-16 network, this modification can extract richer crack detail features while reducing model parameters. Secondly, considering that the Unet network cannot effectively perceive multi-scale crack features in the skip connection stage, a multi-kernel convolution cascade enhancement module is proposed by combining a cascaded connection of multi-kernel convolution groups and multi-expansion rate dilated convolution groups. This module achieves a comprehensive perception of local details and the global content of tunnel lining cracks. In addition, to better weaken the effect of tunnel background clutter interference, a convolutional block attention calculation module is further introduced after the multi-kernel convolution cascade enhancement module, which effectively reduces the false alarm rate of crack recognition. The algorithm is tested on a large number of subway tunnel crack image datasets. The experimental results show that, compared with other crack recognition algorithms based on deep learning, the method in this paper has achieved the best results in terms of accuracy and intersection over union (IoU) indicators, which verifies the method in this paper has better applicability.
ER -