Extracting Surface Defect Contours of Bridge Underwater Pile-pier Structures based on Lightweight Network and Transfer Learning
Underwater pile-pier structures are important components of bridges.Various surface defects occur on these structures due to their complex hydrological environment.Existing methods for the visual detection of such defects have two main issues:① Underwater images are blurred,and the colors are severely distorted;(2)the size of defects cannot be quantitatively identified,and the detection efficiency is low.To solve these problems,this paper proposes a method to extract the contours of underwater pile-pier surface defects by combining an image fusion enhancement algorithm with a deep learning model.First,a pixel-level image fusion algorithm based on point sharpness weights is used,which can fuse two single enhanced images as well as significantly improve image contrast while ensuring effective color correction.Second,the DeepLabv3+semantic segmentation network model is improved in terms of weight,such that the number of weight parameters required for the model can be reduced as much as possible while maintaining the accuracy.Next,an open-source dataset of surface defects in building structures is used to train the backbone feature extraction network layer,and the transfer learning method is applied to the detection task of the object domain.Finally,the image dataset collected from underwater experiments and practical engineering works is used to train the light-weight improved model,establish the underwater pile-pier surface defect contour extraction model,and then verify and test the models.In addition,comparisons focusing on three aspects,namely,comparison with five other commonly used algorithms,comparison of detection results with and without image fusion,and comparison with and without noise effects,are made to verify the robustness and effectiveness of the proposed method.The results show that the image fusion enhancement algorithm proposed in this paper can effectively enhance the detailed features of the defect image contours,and the light-weight improved model has the highest recognition accuracy and can maintain high detection efficiency and robustness.This implies that the proposed method is suitable for the quantitative detection of the surface defect contours of underwater pile-pier structures implanted in small underwater robots for practical bridge structures.
NETL
NSTL
万方数据
