Abstract:

In order to improve the accuracy of bridge crack detection, retain details, and get information on the crack width, the paper proposes a pixel-wise and small sample crack detection method by using U-Net convolutional neural networks. The method uses a U-Net network to extract crack features automatically by using multi-layer convolutions, and uses the superposition of the shallow network and deep network to realize the fusion of local features and abstract features of cracks. This method can retain the details of cracks and greatly improve the accuracy of detection. In order to refine the detection results, the paper presents the threshold method and an improved Dijkstra minimum spanning tree algorithm for eliminating noise and pseudo cracks. Finally, an eight-direction searching method is applied to measure the width of cracks in pixels. Experiments prove that the proposed method can accurately and completely detect bridge cracks and measure the width, which can meet the application requirements.

Key words: image processing, bridge cracks detection, convolutional neural networks, U-Net network