Crack segmentation through deep convolutional neural networks and heterogeneous image fusion

Abstract A DCNN-based crack segmentation methodology is proposed by leveraging heterogeneous image fusion to alleviate image-related disturbances in intensity or range image data and mitigate uncertainties through cross-domain (i.e., intensity and range data domains) feature correlation. Intensity and range images are captured from concrete roadways and integrated through data fusion. Three encoder-decoder networks representing different patterns on exploiting the image data (i.e., fused raw image, raw range image, filtered range image, and raw intensity image) are proposed and compared to benchmarks. Experimental results demonstrate the proposed DCNN exploiting the fused raw image through an “extract-fuse” pattern achieves the most robust and accurate performance on crack segmentation among the implemented DCNNs.