Detecting cracks in concrete structures with the baseline model of the visual characteristics of images

A method based on the baseline model of the visual characteristics of images (BMVCI) is proposed to detect cracks in concrete structures. BMVCI refers to the model, which consists of images of the noncrack areas of a concrete structure with cracks or images of the noncrack state of a concrete structure. Compared with the performance of edge detection (ED) methods for detecting cracks in concrete structures, this baseline model expands the quasi-distance between the edges of cracks and the image background; thus, the crack detection accuracy is effectively improved. Additionally, the discriminative threshold of cracks is quantitatively determined with BMVCI, which avoids the influence of artificial interference when determining the abovementioned threshold used for ED methods. Meanwhile, compared with the methods based on artificial intelligence, such as deep learning (DL), the calculating efficiency of the proposed method is higher because the proposed method converts the high-dimensional image data into low-dimensional digital features for training. With the same small size set of training samples, the accuracy of the crack detection of the proposed method is higher than that of the methods based on the framework of DL. In this study, Gaussian convolution is applied to generate the visual characteristics of images, and then a kernel principal component analysis-based method is implemented to establish the BMVCI. The basic idea of novelty detection is applied to detect cracks in concrete structures. Finally, an experiment on concrete structures is designed and applied to demonstrate the effectiveness of the proposed method.

Automated summary

A method based on the baseline model of the visual characteristics of images (BMVCI) is proposed to detect cracks in concrete structures. Compared with edge detection methods, BMVCI expands the quasi-distance between crack edges and image background, improving crack detection accuracy. Moreover, the method achieves higher computational efficiency and accuracy compared to deep learning methods.