Abstract: Aimed at the drawback of conventional crack detection methods not applicable to the detection of dam surface cracks, this paper presents a new method for quantitative detection based on a full convolution network (FCN). This method combines image preprocessing with morphological post-processing to achieve an improvement on detection accuracy through optimizing raw data and predictions. Oriented at dam surface data, it further improves the accuracy by modifying the traditional FCN network into a more targeted crack detection network, namely a crack full convolution network (C-FCN). And its quantitative information is extracted based on the imaging principle, which avoids complicated camera calibration and is more efficient and objective. We have applied it to in-situ measurements at a dam face and achieved a pixel accuracy of 75.13%, a recall rate of 86.84%, and an intersection ratio of 60.15%. These three indexes are improved by 5.61%, 16.56% and 13.22% respectively in comparison with the traditional FCN network. And the quantified error of detection is less than 5%, and the average opening of the cracks detected is less than 5 mm. Thus, our new detection method would provide a useful tool for dam surface risk assessment and maintenance of water dams.

Key words: deep learning, full convolution neural network, dam surface crack detection, bilateral filtering, quantitative detection