Scholarlink

<p>Urban sewer systems are highly dense and widely distributed. Consequently, sewer defect detection is a frequent, laborious, and time-consuming task. Deep learning techniques rely on a large number of training images for achieving effective recognition. However, when using unevenly distributed sewer defect images, recognition of certain defect types is not satisfactory and the detection process is slow. In this work, a lightweight sewer defect semantic segmentation model based on convolutional block attention module (CBAM) is improved for accurately and automatically recognizing the sewer defects. The improved model uses StyleGAN3 and Deeplabv3+  through the use of MobileNetv2. The experiments are conducted using CCTV images from a field investigation report of drainage pipelines located in Suining, Sichuan province, China. The results show that the improved model has good performance, with a mean pixel accuracy (mPA) of 91.96%, a mean intersection over union (mIoU) of 85.99%, and a mean precision (mPrecision) and overall accuracy (OA) of 92.85% and 96.90%, respectively. Additionally, segmentation speed also increased by 31%. Through data enhancement, networks such as PSPNet, U-Net, and Deeplabv3+  with various backbones show substantial improvements, with mPA, mIoU, mPrecision, and OA increasing by 20% to 30% on average. This demonstrates the effectiveness of StyleGAN3 in recognizing the defects of categories with fewer samples. Furthermore, the inclusion of the CBAM attention module enables the model to focus on the defected edges more effectively, thereby enhancing its pixel-level segmentation capabilities. Consequently, the model accurately segments the indistinguishable cracks in deformations, thus demonstrating precise segmentation effects.</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 novel method for semantic segmentation of sewer defects based on StyleGAN3 and improved Deeplabv3+

  • 期刊论文

  • 发表日期：

    2025年2月24日

查看全文
  • Youlin Li,
  • Yang Yang,
  • Yong Liu,
  • Fengcheng Zhong,
  • Hongrui Zheng,
  • Shiji Wang,
  • Zurui Wang,
  • Zhangyang Huang

摘要

Urban sewer systems are highly dense and widely distributed. Consequently, sewer defect detection is a frequent, laborious, and time-consuming task. Deep learning techniques rely on a large number of training images for achieving effective recognition. However, when using unevenly distributed sewer defect images, recognition of certain defect types is not satisfactory and the detection process is slow. In this work, a lightweight sewer defect semantic segmentation model based on convolutional block attention module (CBAM) is improved for accurately and automatically recognizing the sewer defects. The improved model uses StyleGAN3 and Deeplabv3+  through the use of MobileNetv2. The experiments are conducted using CCTV images from a field investigation report of drainage pipelines located in Suining, Sichuan province, China. The results show that the improved model has good performance, with a mean pixel accuracy (mPA) of 91.96%, a mean intersection over union (mIoU) of 85.99%, and a mean precision (mPrecision) and overall accuracy (OA) of 92.85% and 96.90%, respectively. Additionally, segmentation speed also increased by 31%. Through data enhancement, networks such as PSPNet, U-Net, and Deeplabv3+  with various backbones show substantial improvements, with mPA, mIoU, mPrecision, and OA increasing by 20% to 30% on average. This demonstrates the effectiveness of StyleGAN3 in recognizing the defects of categories with fewer samples. Furthermore, the inclusion of the CBAM attention module enables the model to focus on the defected edges more effectively, thereby enhancing its pixel-level segmentation capabilities. Consequently, the model accurately segments the indistinguishable cracks in deformations, thus demonstrating precise segmentation effects.