Evaluation of MobileNetV3-Large for crack classification across low- and high-resolution images

Introduction This paper evaluates the robustness and generalization ability of five recently developed Convolutional Neural Networks (CNNs), Visual Geometry Group 16 (VGG16), Google Inception Net (GoogLeNet), Mobile Network version 3 Large (MobileNetV3-Large), Efficient Network B0 (EfficientNetB0) and Efficient Network version 2 Small (EfficientNetV2-S), on crack recognition and classification. Methods This study proposes a semantic segmentation based on VGG16- U-Net to address the issue of background noise in the images automatically and the transfer learning with fine-tuning is used to improve the performance of the CNNs in the bridge crack image dataset and building crack image dataset (transverse cracks, vertical cracks, oblique cracks and irregular cracks). Results The results indicate that the MobileNetV3-Large has the best performance. For the low-resolution building crack image dataset, the accuracy of the crack recognition reaches 99.58% and the F1-score reaches 99.60%. The accuracy of the classification reaches 94.70% and the Macro-F1 reaches 94.71%. For the higher resolution bridge crack image dataset, the accuracy of the classification reaches 95.70% and the Macro-F1 reaches 95.67%. Discussion The results show that the MobileNetV3-Large has the best robustness and generalization ability with a small CNN size and the shortest training time.