Computer Vision-Based Bridge Damage Detection Using Deep Convolutional Networks with Expectation Maximum Attention Module.

Abstract

Cracks and exposed steel bars are the main factors that affect the service life of bridges. It is necessary to detect the surface damage during regular bridge inspections. Due to the complex structure of bridges, automatically detecting bridge damage is a challenging task. In the field of crack classification and segmentation, convolutional neural networks have offer advantages, but ordinary networks cannot completely solve the environmental impact problems in reality. To further overcome these problems, in this paper a new algorithm to detect surface damage called EMA-DenseNet is proposed. The main contribution of this article is to redesign the structure of the densely connected convolutional networks (DenseNet) and add the expected maximum attention (EMA) module after the last pooling layer. The EMA module is obviously helpful to the bridge damage feature extraction. Besides, we use a new loss function which considers the connectivity of pixels, it has been proved to be effective in reducing the break point of fracture prediction and improving the accuracy. To train and test the model, we captured many images from multiple bridges located in Zhejiang (China), and then built a dataset of bridge damage images. First, experiments were carried out on an open concrete crack dataset. The mean pixel accuracy (MPA), mean intersection over union (MIoU), precision and frames per second (FPS) of the EMA-DenseNet are 87.42%, 92.59%, 81.97% and 25.4, respectively. Then we also conducted experiments on a more challenging bridge damage dataset, the MIoU, where MPA, precision and FPS were 79.87%, 86.35%, 74.70% and 14.6, respectively. Compared with the current state-of-the-art algorithms, the proposed algorithm is more accurate and robust in bridge damage detection.