A Review of Deep Learning Applications in Concrete Crack Image Recognition and Surface Damage Assessment

Abstract

The morphology and evolution of concrete surface cracks can reflect material deterioration and structural stress conditions, making them important indicators for the inspection of bridges, tunnels, pavements, and building envelopes. In recent years, deep learning methods have been increasingly applied to crack image classification, object detection, semantic segmentation, instance segmentation, and geometric quantification, providing new technical pathways to non-contact inspection and digital maintenance. This review focuses on image-based concrete crack recognition and surface damage assessment, with particular attention to the applicability of convolutional neural networks (CNNs) , transfer learning, one-stage object detectors, Faster R-CNN, Mask R-CNN, U-Net, DeepLab, Transformers, semi-supervised learning, and foundation models to different tasks. Studies have shown that classification models are suitable for rapid screening, object detection models emphasize crack localization, and pixel-level segmentation is more effective for morphological extraction and damage quantification. Current research is shifting from image-level recognition toward precise localization and pixel-level extraction; however, the measurement of actual geometric parameters and quantitative damage assessment remain insufficient. Dataset representativeness, annotation costs, cross-scenario generalization, scale calibration, and engineering validation are the main factors constraining practical application. Future research should strengthen lightweight deployment, annotation-efficient learning, multimodal measurement, and engineering validation to advance crack recognition toward quantitative damage assessment and maintenance decision support.