AN IMAGE-BASED CONCRETE CRACK DETECTION METHOD USING CONVOLUTIONAL NEURAL NETWORKS

This paper proposes a CNN-based crack detection method that can recognize and extract cracks from photos of concrete structures. The algorithm consists of two subsequent procedures, classification, and segmentation, achieved by two convolutional neural networks respectively. First, full images are divided into patches and classified as positive and negative. Then, those sub-images classified as positive are further processed by the image segmentation procedure to obtain the pixel level geometry of the cracks. For the classification part, the performance of transfer learning models based on pre-trained VGG16, Inception V3, MobileNet and DenseNet169 is compared with different classifier. Finally, the CNN based on MobileNet was trained with 30,000 training images and reached 97% testing accuracy and 0.96 F1 score on testing image. For the segmentation part, different neural networks based on the elegant U-net architecture are built and tested. The models are trained with 3840 crack images and annotated ground truth and compared quantitatively and qualitatively. The model with the best performance reached 88% sensitivity on test data set. The combination of the classification and segmentation neural networks achieves an image-based crack detection method with high efficiency and accuracy. The algorithm can process any full image size as input. Compared with most machine learning based crack detection algorithms using sub-image classification, a relatively larger patch size is used in this paper and in this way the classification is more robust and accurate. On the other hand, the negative areas in the full image will not be concerned in the segmentation procedure and this fact not only saves a lot of computational power but also significantly increases the accuracy compared to the segmentation performed on full images.