Scholarlink

The advent of new technologies and the necessity for an alternative to manual inspection have led to the usage of computer vision techniques and deep learning methods for structural inspection. The current study aims to identify efficient deep CNN architectures, previously developed for image classification and tested on benchmark datasets like ImageNet. The primary focus of this study is binary classification, i.e., the model predicts the presence of damage in a given image. To achieve the research objectives, two architectures, Resnet and DenseNet, and their variants were compared. The performance of models was compared with respect to precision, recall, and F1 scores. The data for training and validation was formed from merging existing datasets that contained three types of damages, namely cracks, spalling and rebar exposure. For testing the model’s performance, inference was done on a new dataset. Further, visual explanations for the models were obtained using the Grad-CAM method to understand the localization ability of the model.

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

Evaluation of Deep Learning Architectures for Multiple Damage Detection on Concrete Surface

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Lokeswari Malepati,
  • Nikesh Thammishetti,
  • Nagarajan Ganapathy,
  • S. Suriya Prakash,
  • Vedhus Hoskere

摘要

The advent of new technologies and the necessity for an alternative to manual inspection have led to the usage of computer vision techniques and deep learning methods for structural inspection. The current study aims to identify efficient deep CNN architectures, previously developed for image classification and tested on benchmark datasets like ImageNet. The primary focus of this study is binary classification, i.e., the model predicts the presence of damage in a given image. To achieve the research objectives, two architectures, Resnet and DenseNet, and their variants were compared. The performance of models was compared with respect to precision, recall, and F1 scores. The data for training and validation was formed from merging existing datasets that contained three types of damages, namely cracks, spalling and rebar exposure. For testing the model’s performance, inference was done on a new dataset. Further, visual explanations for the models were obtained using the Grad-CAM method to understand the localization ability of the model.