Abstract
This paper proposes a novel crack detection method for concrete structures based on Fourier image enhancement and convolutional neural network (CNN). Cracking is regarded as an important indication for structure aging and durability decline, so the detection of cracks becomes more and more important. The CNN is used to detect the existence of cracks automatically. The original crack images may be disturbed by many factors, such as image blur, distortion and so on. In order to improve crack detection accuracy, it is necessary to pre-process the original images. This paper introduces a frequency-domain enhancement algorithm based on the Fourier transform, which is effective for crack detection from low-quality crack images (such as images with blurs and distortion). The results show that the testing accuracy of the crack images after Fourier enhancement reaches 100%. In this paper, a control experiment is designed to illustrate the effectiveness of this method, in which the crack images are not pre-processed and the testing accuracy is only 87.5%, and when the crack image is processed by median filter, and the testing accuracy is only 91.67%. The experimental results show that Fourier enhanced crack images can effectively improve the accuracy of crack detection and make the CNN training faster.