A channel attention based deep neural network for automatic metallic corrosion detection. (October 2021)

• Record Type: Journal Article
• Authors: Zhang, Sheng; Deng, Xinling; Lu, Yumin; Hong, Shaozheng; Kong, Zhengyi; Peng, Yongli; Luo, Ye
• Is Part Of: Journal of building engineering. Volume 42(2021)
• Publication Date: 2021-10
• Subjects: Deep learning -- Convolutional neural network -- Metallic corrosion detection -- Channel attention -- Residual neural network; Building -- Periodicals; 690.05
• DOI: 10.1016/j.jobe.2021.103046
• Languages: English
• ISSNs: 2352-7102

Abstract

Abstract: Metallic corrosion in civil infrastructure systems has incurred astronomical costs and considerable risks to industrial communities worldwide. Most existing research focuses on manually extracted features or employing complex convolutional neural networks for powerful deep feature learning on images of metal sheets. But few of them pay attention to learning discriminative deep features for metallic corrosion detection. In this paper, we propose a Channel Attention based Metallic Corrosion Detection method (CAMCD), by which the corroded regions with multiple distinct levels can be automatically detected patch-wisely. Correspondingly, to learn the patch-wise features and discriminate them among various corrosion levels, a CAMCD network is built by embedding SE blocks into the deep residual network; thus, the important features of various corroded regions are highlighted by weighting with the learned weights of channel attentions. Experimental results on our collected metallic corrosion dataset validate the superiority of our proposed CAMCD method over other existing approaches on corroded region detection. And the visualizations of the feature maps weighted by the channel attention further confirm the effectiveness of our CAMCD network on discriminative feature learning. Highlights: A Channel Attention based Metallic Corrosion Detection method is proposed to detect various levels of corroded regions. SE-ResNet bottlenecks are employed in CAMCD to learn weights and discriminative features. Experimental results validate the superiority of our method over other existing corrosion detection methods.