[IPOL announce] new article: An 'All Terrain' Crack Detector Obtained by Deep Learning on Available Databases

[announcements about the IPOL journal]

A new article is available in IPOL: http://www.ipol.im/pub/art/2020/282/

Sébastien Drouyer,
An 'All Terrain' Crack Detector Obtained by Deep Learning on Available 
Databases,
Image Processing On Line, 10 (2020), pp. 105–123.
https://doi.org/10.5201/ipol.2020.282

Abstract
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.