E³-Net: Event-Guided Edge-Enhancement Network for UAV-based Crack Detection

Buildings, bridges, and other exteriors of infrastructure are commonly susceptible to damages that may reduce safety and longevity. When left unattended, small cracks can cause catastrophic failures. However, crack inspection is a challenging process that requires significant human effort and can expose inspectors to risks. To address this challenge, the use of Unmanned Aerial Vehicles (UAVs) has emerged as an approach to survey areas that are difficult to reach. However, the cracks are small and may blend into background textures, making them difficult to discriminate. Furthermore, the motion of quadrotors can potentially cause motion blur, increasing the difficulty of crack detection. To address these challenges, we propose a neural network designed to detect cracks robustly. Our neural network architecture includes two stages. First, a V2E (Video-to-Event) model transforms the image data into Dynamic Vision Sensor (DVS) events. This procedure aims to make the output crack feature identifiable in the presence of motion blur. Then, an edge-enhancement net that combines UNet and non-local block net is designed to augment the crack feature and suppress the background texture concurrently. The background texture suppression is achieved by training on our annotated crack dataset using a triplet loss. Lastly, these proposed cracks can be highlighted by the YOLOv8n-seg network. Experimental results demonstrate that the proposed network can effectively detect small wall cracks in the presence of diverse background patterns.