Intelligent identification method of apparent defects in underwater concrete structures based on deep learning(PDF)
《建筑科学与工程学报》[ISSN:1673-2049/CN:61-1442/TU]
- Issue:
- 2026年01期
- Page:
- 67-74
- Research Field:
- 智能检测与建造技术专栏
- Publishing date:
Info
- Title:
- Intelligent identification method of apparent defects in underwater concrete structures based on deep learning
- Author(s):
- CHEN Dejin; ZHANG Donglin; YE Xijun; ZHENG Wenzhi; ZHOU Junyong
- School of Civil Engineering and Transportation, Guangzhou University, Guangzhou 510006, Guangdong, China
- Keywords:
- underwater concrete; YOLOv9; image fusion; twostage recognition
- PACS:
- -
- DOI:
- -
- Abstract:
- In underwater environment, the scattering and absorption of water media and the influence of water flow, it is easy to cause blurring, atomization, distortion and low resolution of underwater images, which subsequently reduces the accuracy of automated structural defect identification based on deep learning. Through twostage processing, a model integrating image fusion algorithm and target recognition algorithm was established. In stage 1, the image fusion algorithm was used to enhance the detectability of defects. The algorithm achieved feature image weight fusion using an improved CycleGAN and a multiscale retinex algorithm with color protection (MSRCP). The improvements to CycleGAN included three parts, which were optimizing the cycle consistency loss, replacing the activation function, and incorporating a new attention mechanism. These enhancements could effectively enhance the image feature extraction capability, facilitate a more efficient training process, and reduce model weights. Stage 2 employed the YOLOv9 object recognition algorithm to achieve rapid and accurate identification and classification of underwater structural surface defect types in complex water environment. The results show that by enhancing underwater structural surface defect images, the mAP0.5 identification rate of the proposed method reaches 74.3%, which is 11.6% higher than that of the untreated case. The recognition accuracy of the proposed method is 74.5 %, 58.5 % and 90.0 % for the recognition of three classic apparent diseases of “concrete spall”, “exposed rebar”, and “concrete cracking”, and compared with the untreated case, it is increased by 1.9 %, 5.6 % and 26.1 %, respectively.
Last Update: 2026-01-20