Enhancing Concrete and Pavement Crack Prediction through Hierarchical Feature Integration with VGG16 and Triple Classifier Ensemble

When assessing the structural integrity of concrete infrastructure, detecting fractures in roads is paramount. However, conventional image-based algorithms often encounter challenges with noisy concrete surfaces, particularly in real-world scenarios like autonomous car road recognition. To address this, traditional techniques often require intricate preprocessing, which can be inconvenient in noisy and varied environments. This paper introduces a new image-based crack detection approach, the Majority Voting Crack Detection method, aimed at avoiding these hurdles. This method leverages a deep convolutional neural network-CNN in organization with three machine learning classifiers: Random Forest, Logistic Regression, and Support Vector Machine. By combining these techniques, it offers a robust solution for identifying fracture features even amidst noisy and heterogeneous conditions. The research begins by training the three state-of-the-art classifiers using a dataset of 30,000 images. Subsequently, a majority voting scheme is implemented, harnessing the collective strengths of VGG16 and the three prominent machine learning classifiers to enhance accuracy. Through systematic comparison, an optimal base learning rate of 0.001 is determined, resulting in a highest validation accuracy of 100%. Subsequently, during the testing phase, this optimal learning rate is applied. A set of six thousand test images, each with a resolution of 224 × 224 pixels and distinct from the training or validation datasets, is utilized to evaluate the robustness and adaptability of the trained models. The findings underscore the practical utility of this proposed approach by demonstrating its effectiveness in accurately identifying road fractures in images captured on real concrete surfaces.