Edge-AI Based Real-Time Crack Detection and Severity Classification for Concrete Bridge Inspection Using Deep Convolutional Networks

Routine visual inspection of concrete bridges is labour-intensive, subjective and increasingly impractical given India's expanding network of over 175 000 road bridges. This paper presents an edge-AI inspection platform that performs real-time crack detection, segmentation and severity classification onboard an unmanned aerial vehicle. A lightweight encoder–decoder network combining a MobileNet-V3 backbone with a U-Net decoder was trained on a curated dataset of 11 240 bridge surface images and deployed on an NVIDIA Jetson Orin Nano edge device. The proposed model achieved a mean intersection-over-union of 0.86 and a sustained inference rate of 22 frames per second at 512×512 input resolution, outperforming three baseline architectures while remaining within a 7 W power envelope. Field validation across two bridge structures in Karnataka demonstrated reliable detection of hairline through severe crack categories with geo-tagged outputs streamed to a cloud dashboard for engineering review