CrackSentinel-MobileNet: A Lightweight Transfer Learning Framework for Automated Structural Surface Crack Detection in Civil Infrastructure
Cracks in concrete can be perceived to be minor initially, just being a thin dark line in a bridge support pillar or a retaining wall. However, these cracks are hard to ignore once a few seasons of monsoons come and go. Water enters the crack, the reinforcement bar starts to rust, the strength of the structure reduces, and the initial minor problem evolves into a major issue. The important thing here is to identify this trend, and that’s exactly where the role of an inspection system comes into play. Currently, manual inspections do not have enough accuracy for this task.Inspectors are expensive, their schedules stretch, and two people looking at the same wall photograph do not always reach the same conclusion [1]. This paper describes CrackSentinel-MobileNet, a system we put together and tested for exactly this kind of binary screening task: given a surface photograph, is there cracking present or not? The approach takes MobileNetV2—a compact network trained on ImageNet — locks its weights in place, and trains a small custom head on top of it. That head is: global average pooling, batch normalisation, a 128unit dense layer, half-rate dropout, then a sigmoid. Training used 31,999 photographs over five epochs; 7,999 separate images were held back for testing and never touched during training. Results on that test set: 99.82% accuracy, 99.92% recall, $\mathrm{AUC}=1.000$. Out of 7,999 images, 14 came out with the wrong label. That is a number we feel comfortable defending in a real inspection screening context.