Efficient hybrid ensembles of CNNs and transfer learning models for bridge deck image-based crack detection

Automatic image-based crack detection of concrete bridge decks contributes to safer bridge operation and bridge health monitoring. Existing models suffer from overfitting and low generalization abilities. Moreover, their performances highly depend on the model architecture, training method, data source, etc. To address these challenges, several hybrid self-designed and transfer learning ensemble models have been introduced for the efficient and accurate intelligent crack detection. Firstly, some self-designed convolutional neural networks (CNNs) are constructed from scratch using labeled crack and non-crack images from modified existing bridge deck image dataset. Secondly, some pretrained transfer learning models, namely the VGG16, VGG19, ResNet50, MobileNetV3Small Model, InceptionResNetV2, EfficientNetV2B0, Xception, and InceptionV3 are adopted to check the efficiency of transfer learning in detecting cracks in bridge decks images. Using the developed CNNs and transfer learning models, several ensemble learning models between the self-designed CNNs, transfer learning CNNs, as well as hybrid self-designed CNNs and transfer learning models are developed. The ensemble learning strategies including the weighted average, stacking, Adaboost, Gradient boosting, and XGBoost ensembles are utilized to construct the ensemble learning models aiming to increase the prediction accuracy and improve generalization ability. Results indicate that the hybrid ensemble learning between the self-designed CNNs and the transfer learning models highly improve the precision and accuracy of the individual models and can be well implemented for image-based bridge deck crack detection.