DAMFE-Net: A Few-shot Crack Segmentation Model Based on Transfer Learning for Earth Dams

Some universal segmentation models have already shown good performance on public datasets, but they perform poorly in detecting cracks in earth dams. This deficiency stems from two aspects: one is the shortage of the dataset about cracks on earth dams, and the other is the complex and unstable appearance influenced by material and environmental conditions, making crack segmentation more susceptible to background noise. In this paper, we put forward a crack segmentation model based on transfer learning for earth dams with higher accuracy in few-shot conditions, named Dual Attention and Multi-scale Feature Enhancement Network (DAMFE-Net). It introduces a Spatial and Channel Squeeze and Excitation module (SCSE) to electively enhance key features for better extraction, and it incorporates the Multi-scale Feature Enhancement Module (MFEM) suppresses irrelevant features, and enhances crack feature extraction in earth dam images, reducing high background noise. The MFEM learns the weight of each scale feature map and fuses the semantic information of features at different levels, so it improves the representation of effective features. The F1-score for crack segmentation on earth dams with our model is 86.29%, surpassing DeepLab v3+ by 16.4%, which indicates that the proposed model can accurately segment cracks on earth dams.