Performance Comparison of Multiple Convolutional Neural Networks for Concrete Defects Classification.

No prior work had resolved the limitations of deep learning models in identifying multiple simultaneous concrete surface issues. Prior research has shown that vision-based systems facilitate structural health monitoring through automated pattern recognition. That uncertainty drove the need to expand beyond single-defect detection frameworks. Most existing approaches focus exclusively on one specific type of damage. This gap motivated the development of a more versatile classification system. It was already known that artificial intelligence improves inspection efficiency. However, the performance of various architectures on diverse concrete damage remains poorly understood. This study addresses the lack of comparative benchmarks for multi-class concrete defect identification.

The aim of this study is to evaluate the performance of various deep learning models in classifying concrete surface defects. The researchers sought to address the limitations of existing methods that typically detect only one damage type. They developed a unique, non-augmented dataset containing original images of cracks and spalling. This effort was motivated by the need for more versatile and accurate structural health monitoring systems. The authors intended to compare five pre-built architectures to identify the most reliable solution. They also aimed to determine the best hyper-parameter configurations for these models. By conducting a detailed sensitivity analysis, the team explored how different optimizers and learning rates influence classification outcomes. This work provides a benchmark for selecting optimal image recognition tools in civil engineering applications.

Review Approach involved applying five distinct pre-built architectures to a newly curated image collection. The researchers focused on classifying two specific damage categories without using standard data expansion techniques. They performed a rigorous sensitivity analysis to evaluate different training configurations. This process included testing various optimizers and learning rate settings to refine model behavior. The team compared the output of VGG-19, ResNet-50, InceptionV3, Xception, and MobileNetV2 systematically. They prioritized maintaining original image integrity to simulate authentic field conditions. Each model underwent evaluation to determine its classification proficiency for cracks and spalling. This structured comparison allowed the authors to identify the most effective parameters for automated inspection tasks.

Key Findings From the Literature indicate that InceptionV3 outperformed all other models with 91% accuracy. The analysis revealed a precision of 83% and a recall of 100% for this specific architecture. These metrics confirm the model's high proficiency in identifying multiple concrete damage types. The researchers found that stochastic gradient descent served as the optimal optimizer for the top-performing system. A learning rate of 0.001 proved most effective during the training phase. Other tested models failed to match these specific performance benchmarks. The study highlights the success of InceptionV3 in handling complex classification tasks without relying on augmented data. These results establish a clear hierarchy of model effectiveness for structural health monitoring applications.

Synthesis and Implications suggest that InceptionV3 provides superior classification accuracy for concrete surface damage compared to other tested architectures. The authors propose that selecting the correct optimizer and learning rate significantly influences model success. Their findings indicate that stochastic gradient descent paired with a 0.001 learning rate optimizes detection outcomes. The researchers conclude that avoiding image augmentation preserves real-world conditions for training datasets. This study demonstrates that deep learning models possess high proficiency in recognizing multiple distinct defect types. The evidence supports the use of pre-built models for enhancing structural health monitoring workflows. These results highlight the importance of hyper-parameter sensitivity analysis in achieving peak classification performance. The authors maintain that their unique, large-scale dataset provides a robust foundation for future automated inspection systems.