Pixel-Level crack detection using an attention mechanism

Abstract

This paper proposes a pixel-by-pixel automatic crack detection method ECCrack. The method based on the combination of a channel attention mechanism and fully convolutional neural network. The codec structure is combined with the channel attention mechanism to enhance network’s ability to utilize crack feature information. In the experiment, we verified this method on the road crack data set, and the F1-score reached 92.83, which is better than other methods. Besides, we also conducted ablation experiments on this method to prove the effectiveness of the increased mechanism.