Adaptive and Explainable Deep Learning-Based Rapid Identification of Architectural Cracks

Concrete architectural structures are widely used in urban construction, making the health diagnosis and maintenance of these structures increasingly essential and urgent. Crack identification is crucial for maintaining the structural integrity and safety of buildings. Traditional methods rely on manual inspection, which is plagued by low accuracy, inefficiency, and safety hazards. This paper proposes a technique combining an attention-based SqueezeNet network with Gradient-weighted Class Activation Mapping (Grad-CAM) for automatically recognizing and visually explaining building cracks. By integrating the Squeeze-and-Excitation (SE) attention mechanism with the lightweight SqueezeNet network, this method can adaptively adjust the importance of feature channels by learning global information, effectively improving the network’s accuracy and efficiency. The experimental results show that the Att-SqueezeNet model achieved a high precision of 0.995, a training time of only 133 seconds, and a model size of 4.9M, significantly outperforming models such as SqueezeNet, RF, CNN, VGG-19, and B-CNN. This demonstrates its robustness, rapid identification and suitability for practical applications and building crack identification. Moreover, the utilization of Grad-CAM for visualization not only offers an intuitive explanation of the model’s decision-making process but also provides a more comprehensible understanding of crack detection results. This is crucial for advancing building maintenance automation, reducing reliance on manual labor, and increasing the precision and reliability of detection tasks.