LightAUNet: A Lightweight Fusing Attention Based UNet for Crack Detection

Abstract

As a common pavement defects, crack Detection plays an important role in ensuring traffic safety. Thus accurate and fast detection of pavement cracks has important research value. At present, Attention based UNet models have achieved promising results by using attention block to suppress interference. However, these models have two defects. Firstly most of the existing attention mechanisms, only consider selecting spatial features, without considering selecting channel features. Secondly the size of these models is very large, which could not be computed in mobile devices in real time. To solve the problem, we proposed a Lightweight Fusing Attention based UNet (LightAUNet) model. In the LightAUNet, in order to reducing the high complexity of UNet based models, the feature extraction network is redesigned by us, that is the Depth-Separable Convolution is introduced to replace the standard convolution. In addition, features are designed to be selected and fused from two dimensions, spatial dimension and channel dimension. Therefore, we could get more comprehensive crack fetaures by fusing these two features, and thus it could surppress more interference than traditional attention block. Compared with the state-of-the-art crack detection method, our proposed LightAUNet shows competitive segmentation performance. Experimental results show that this model could achieve accuracy of 91%, 81%, 98.2% on Cracktree200, CRACK500 and CFD dataset respectively, which is 2%, 5% and 1% higher than that of the traditional attention based UNet model, while the size of our model is only 7M.