Quantitative Performance Evaluation of Improved U-Net Model in Crack Detection

We enhanced crack detection by combining channel and spatial attention mechanisms to overcome the limitations of the U-Net model. Detecting cracks is pivotal for ensuring infrastructure and household safety. However, challenges persist in preserving image details and extracting meaningful features especially due to the scarcity of high-resolution data. We investigated the impact of channel and spatial attention on crack detection using a publicly available dataset for evaluation. The proposed method augmented the U-Net model with attention mechanisms by integrating channel and spatial attention to discern inter-channel correlations and capture intricate details. The quantitative analysis results demonstrated significant improvements over the baseline U-Net model. The refined model enhanced accuracy, recall, and F1 score, showing the effectiveness of the combined attention mechanisms. Specifically, channel attention is to understand inter-channel correlations, while spatial attention is used to capture fine spatial details. The two allow for a synergistic enhancement in crack detection. The result of this research advances crack detection methodologies based on channel and spatial attention mechanisms. These mechanisms signify safer infrastructure and improved household safety.