<p>It is of great significance to detect the cracks of the pavement accurately for safety. However, the complex background and random shape of cracks in real pavement crack images limit the development of crack detection. To address this challenge, we propose a novel cement pavement crack detection network. Firstly, we propose a new backbone network named EPSA-DS-CrackNet, which uses efficient pyramid squeeze attention (EPSA) blocks to extract crack features of different sizes, and adds deep supervision (DS) behind them to aggregate multi-level features and obtain the final prediction map. Secondly, a new hybrid spatial pyramid pooling (HSPP) module is designed by us, which carries out cross-latitude interaction from global, local, vertical and horizontal dimensions to enhance crack detailed features and eliminate redundances. Finally, in order to recover the crack location information, we design a multiple multi-layer fusion (MMF) method to establish a simple and efficient cross-scale connection between feature maps of different sizes. Evaluations of DeepCrack, CRACK500 and EdmCrack600 public datasets show that our approach outperforms other new approaches, which achieves F-score values of 87.3<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation>, 75.8<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation>, 62.0<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation> and <i>MIoU</i> values of 88.2<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation>, 79.1<InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation>, 72.1<InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation>. The code is available at: <a href="https://github.com/zwl228/EPSA-DS-CrackNet/tree/main">https://github.com/zwl228/EPSA-DS-CrackNet/tree/main</a>.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

A crack detection algorithm with multiple multi-layer fusion and hybrid spatial pyramid pooling for cement pavement

  • Zhong Qu,
  • Wenli Zhang

摘要

It is of great significance to detect the cracks of the pavement accurately for safety. However, the complex background and random shape of cracks in real pavement crack images limit the development of crack detection. To address this challenge, we propose a novel cement pavement crack detection network. Firstly, we propose a new backbone network named EPSA-DS-CrackNet, which uses efficient pyramid squeeze attention (EPSA) blocks to extract crack features of different sizes, and adds deep supervision (DS) behind them to aggregate multi-level features and obtain the final prediction map. Secondly, a new hybrid spatial pyramid pooling (HSPP) module is designed by us, which carries out cross-latitude interaction from global, local, vertical and horizontal dimensions to enhance crack detailed features and eliminate redundances. Finally, in order to recover the crack location information, we design a multiple multi-layer fusion (MMF) method to establish a simple and efficient cross-scale connection between feature maps of different sizes. Evaluations of DeepCrack, CRACK500 and EdmCrack600 public datasets show that our approach outperforms other new approaches, which achieves F-score values of 87.3 \(\%\) , 75.8 \(\%\) , 62.0 \(\%\) and MIoU values of 88.2 \(\%\) , 79.1 \(\%\) , 72.1 \(\%\) . The code is available at: https://github.com/zwl228/EPSA-DS-CrackNet/tree/main.