Res-FPN–SqueezeNet Segmentation Architecture for Enhanced Identification of Concrete Cracks within a Deep-Learning Framework

To enhance the precision and efficiency of concrete crack detection and segmentation in concrete structures, this paper proposes a novel Res-FPN–SqueezeNet segmentation model improved by the SqueezeNet convolutional neural network based on a deep learning framework. The innovation lies in integrating features from residual networks (ResNets) and feature pyramid networks (FPNs) for multiscale feature extraction, significantly boosting the model’s ability to capture fine crack details while preserving a lightweight architecture ideal for real-time applications. The enhanced model was trained on a combination of an open-source concrete crack image data set and laboratory-collected samples, ensuring robust performance across diverse conditions. Comparative evaluations with state-of-the-art models and traditional segmentation techniques demonstrated that the Res-FPN–SqueezeNet model achieved a 1.76% increase in precision over the original SqueezeNet, along with superior geometric accuracy. Crack length and area detection errors remained below 5% and width errors under 10%, validating the model’s ability to handle various crack types. Those results underscored the model’s potential for practical application in the automated monitoring and maintenance of concrete infrastructures, offering a scalable and efficient solution to key challenges in structural integrity assessment.