Attention enhanced EfficientNet for concrete structure crack classification with generative adversarial network augmented data
- 3 March 2025
- journal article
- Published by Wiley in Structural Concrete
- Vol. 26 (4) , 4651-4666
Abstract
Cracks are an important indicator of the decline in the load‐bearing capacity of buildings. Therefore, it is of great significance to detect and classify cracks in reinforced concrete (RC) building exterior walls. Accurately and automatically classifying cracks remains challenging due to the highly irregular nature of crack images, irregular lighting conditions, and background texture noise. An EfficientNet network model combined with a HiLo attention mechanism was proposed to achieve precise identification and classification of RC exterior wall cracks. Firstly, existing crack datasets were categorized, and a combination of classical data augmentation and conditional generative adversarial networks was used to augment the data, improving the model's generalization ability under different imaging conditions and mitigating the adverse effects of the unbalanced dataset. Furthermore, crack images were divided into high‐frequency (Hi‐Fi) and low‐frequency (Lo‐Fi) feature maps by applying the HiLo attention mechanism. Hi‐Fi module suppressed background noise and captured the edge details of cracks by retaining relatively high‐resolution feature maps. Lo‐Fi module extracts global features of cracks through window segmentation and average pooling operations, thereby enhancing the classification capability of model. The experimental results showed that the HiLo‐EfficientNet model achieved the best image classification performance with an overall classification accuracy of 91.63% compared to the original and mainstream deep learning models.Keywords
Funding Information
- National Natural Science Foundation of China (U22A20229)
- Natural Science Foundation of Shandong Province (ZR202103020358)
This publication has 33 references indexed in Scilit:
- Study on the use of different machine learning techniques for prediction of concrete properties from their mixture proportions with their deterministic and robust optimisationAI in Civil Engineering, 2024
- The use of generative adversarial networks in medical image augmentationNeural Computing and Applications, 2023
- Research on improving the durability performance of damaged concrete based on ultra high performance concrete repairEngineering Reports, 2023
- Dynamic weight HiLo attention network for medical image multiple organ segmentationInternational Journal of Imaging Systems and Technology, 2023
- Assessing compressive strengths of mortar and concrete from digital images by machine learning techniquesFrontiers of Structural and Civil Engineering, 2022
- Structural crack detection using deep convolutional neural networksAutomation in Construction, 2022
- Curve‐based crack detection using crack information gainStructural Control and Health Monitoring, 2021
- Dataset retrieval system based on automation of data preparation with dataset description modelConcurrency and Computation: Practice and Experience, 2019
- Crack detection using image processing: A critical review and analysisAlexandria Engineering Journal, 2018
- Crack detection and characterization techniques-An overviewStructural Control and Health Monitoring, 2014