Comparative Study of Lightweight CNNs for Concrete Crack Detection with GPU-Based Inference

Concrete crack detection is a critical task in structural health monitoring and maintenance planning, particularly in aging infrastructures. While deep learning has significantly advanced the automation of crack identification, deploying such models in real-world scenarios requires balancing detection accuracy with computational efficiency—especially on resource-constrained hardware. This study conducts a comparative evaluation of four state-of-the-art lightweight convolutional neural network (CNN) architectures-MobileNetV۲, ShuffleNet V۲, Efficient Net-B۰ and Squeeze Net-for the binary classification of surface cracks in concrete. The primary aim is to assess the performance of these models in terms of both predictive accuracy and computational practicality when trained and inferred on a consumer-grade GPU (NVIDIA GeForce GTX ۱۰۵۰ Ti) using the SDNET۲۰۱۸ benchmark dataset. The research method involves transfer learning with fine-tuning, implemented in TensorFlow/Keras, followed by a thorough evaluation using standard metrics including Accuracy, Precision, Recall, F۱-score, Intersection over Union (IoU), and mean Average Precision (mAP). Each model's training and validation performance over epochs is analyzed through loss and metric progression graphs. Test performance is quantitatively compared in a comprehensive results table, supplemented by visual examples of crack localization from the highest-performing model. Bar charts and radar plots are used for intuitive cross-model comparison. The findings demonstrate that while all evaluated models show promise for practical deployment, significant trade-offs exist between detection performance and model complexity. The study offers actionable insights for researchers and engineers seeking efficient yet reliable CNN-based solutions for concrete surface inspection, especially in environments with limited computational resources. This work contributes to the advancement of lightweight AI models in civil infrastructure monitoring and lays the groundwork for further real-time or edge-device deployments.

کلیدواژه ها:

Lightweight CNNs ، Concrete Crack Detection ، GPU-Based Inference ، SDNET۲۰۱۸