Multidamage Identification in High-Resolution Concrete Bridge Component Imagery Based on Deep Learning

Abstract

The identification of damage to concrete bridge surfaces is of great significance to maintaining the durability and reliability of bridges. However, it is difficult to identify small areas of damage, especially in high-resolution images. To conduct damage identification research in a targeted manner, this paper proposes a method of multidamage identification in high-resolution images based on You Only Look Once version 5 (YOLOv5). In this paper, a data set labeled with four types of damage (crack, spallation, hole, and rebar) is used for training, validation, and testing. To ensure that the network adequately learns the damage features, the high-resolution images are cropped into subimages via the autoadaptive window cropping method (AWCM) proposed in this paper. The cropping method can crop images according to the label information and protect the damage features from destruction during the cropping process. To avoid overfitting, it is necessary to balance the volume of different types of damage. After balancing, the count for each category is as follows: crack (4,980), spallation (5,225), hole (5,211) and rebar (5,020). The balanced data set can be used to train the deep learning network and construct the multidamage identification model. After identification, the subimages, and the prediction boxes of damages in them are restored to their original high-resolution images. The results show that the mean average precision (mAP) of all classes is 94.2%, and the values for cracks, spallation, holes, and rebars are 86.9%, 98.1%, 92.3%, and 99.4%, respectively, which indicates that the proposed method outperforms the other three methods (inputting original images directly, the sliding window cropping method, and the random centroid cropping method).