UMDA: Lightweight and Efficient Crack Segmentation Model

Efficient crack detection is vital for infrastructure safety, yet many deep learning models sacrifice practicality for precision, demanding resources beyond the reach of field-deployable devices. This paper presents UMDA (U-MobileNetV3-DECA-AUX), a lightweight crack segmentation model that balances accuracy and efficiency. Built on MobileNetV3-large, UMDA uses bilinear upsampling for smooth outputs, a Dilated Efficient Channel Attention (DECA) module to enhance feature focus, and an auxiliary head to improve performance. On the DeepCrack data set, UMDA achieves a Dice coefficient of 82.5% and an IOU of 75.3%, processing 480×480 images in just 16.34 ms—outpacing traditional models. Integrated with unmanned aerial vehicle (UAV) technology, it excels in segmenting fine bridge cracks, boosting both safety and efficiency. UMDA offers a scalable solution for real-time crack monitoring across diverse engineering contexts.