Dam Crack Image Detection Model on Feature Enhancement and Attention Mechanism

Abstract

Dam crack detection can effectively avoid safety accidents of dams. To solve the problem that the dam crack image samples are not available and the traditional algorithm detects cracks with low accuracy, we provide a dam crack image detection model based on crack feature enhancement and attention mechanism. Firstly, we expand the dam crack image dataset through a generative adversarial network based on crack feature enhancement (Cracks Enhancements GAN, CE-GAN). It can fully expand the dam crack data samples and improve the quality of the training data. Secondly, we propose a crack image detection model based on the attention mechanism (Attention-based Faster-RCNN, AF-RCNN). The attention mechanism is added in the crack detection module to give different weights to the proposal boxes around the crack target and fuse the candidate boxes with high weights to accurately detect the crack target location. The experimental results show that our algorithm achieves 81.07% mAP on the expanded dam crack dataset, which is 8.39% higher than the original Faster-RCNN algorithm. The detection accuracy is significantly improved compared with other traditional dam crack detection algorithm models.