Efficient Channel Attention U-Net For Mesh Crack Detection

Abstract

Cracks are an extremely common disease in concrete structures, and effective detection can prevent greater damage. Although the existing deep learning-based crack detection technology is well developed, there are still some shortcomings, such as the lack of detection of mesh cracks due to discontinuity of extracted lines. In this paper, a new deep learning model based on U-Net network is proposed, aiming to improve the ability of detecting mesh cracks by increasing the continuity of lines. We enhance the extraction of semantic information of fine branch cracks by introducing the efficient channel attention (ECA) blocks, which replaces the simple copy between the contracting path and the corresponding expansive path, and makes the network focus more on thin crack pixels and suppress the pixels of the background. The results show that our model has reached the state-of-the-art performance on our dataset.