Road Crack Detection Using Quaternion Neural Networks

Automatic road crack detection can play an significant role in improving driving safety. However, a number of factors make this task challenging. Background complexity, as well as the fact that cracks are visually inhomogenous to one another also contribute to this difficulty. Furthermore, cracks in road surfaces can be easily confused with foreign objects, shadows and background textures leading to detection ambiguities. Timely detection of cracks is important for both drivers and road maintenance crews. Quaternion neural networks (QNNs) are a relatively new class of neural networks that employ quaternion-valued activations and parameters. They benefit from reduced costs in terms of hardware as they require fewer parameters. In this work, we explore the usability of QNNs for automatic road crack detection. To this end, we propose quaternionic versions of deep networks and evaluate their performance in datasets with images of road cracks. We show that the proposed models are light-weight in terms of parameter requirements while they are on par in terms of performance with real-valued networks for crack detection, with potential applications in resource constrained scenarios or in cases where few training data is available.