An "All Terrain" Crack Detector obtained by Deep Learning on Available Databases

We present a general deep learning method for detecting cracks on all sorts of surfaces. For making this method robust to different types of cracks and acquisition procedures, we have trained our method on four datasets - Crack500, DeepCrack, SDNet2018 and CrackForest. We have also labelled a part of the SDNet2018 dataset so that it contains semantic labels, as it originally only proposed crack/non-crack classifications on the image level. To validate our approach, we perform a cross-dataset study where we train the model on a subset of the datasets and test it on another subset. Results of this study show that training the model on these various datasets makes it more robust to new images, outperforming existing classical and deep learning methods. In order to make our method even more robust to different objects, scenes and illuminations, we have also added images from the Flickr website, leading to an important drop in false positives on extra dataset images. The network seems to function well on images not belonging to any of the datasets, and its publication in IPOL will allow users to enrich further training.