Convolutional neural networks for steel surface defect detection from photometric stereo images

Abstract

Convolutional neural networks (CNNs) achieved impressive recognition rates in image classification tasks recently. In order to exploit those capabilities, we trained CNNs on a database of photometric stereo images of metal surface defects, i.e. rail defects. Those defects are cavities in the rail surface and are indication for further surface degradation right up to rail break. Due to security issues, defects have to be recognized early in order to take countermeasures in time. By means of differently colored light-sources illuminating the rail surfaces from different and constant directions, those cavities are made visible in a photometric dark-field setup. So far, a model-based approach has been used for image classification, which expressed the expected reflection properties of surface defects in contrast to non-defects. In this work, we experimented with classical CNNs trained in pure supervised manner and also explored the impact of regularization methods such as unsupervised layer-wise pre-training and training data-set augmentation. The classical CNN already distinctly outperforms the model-based approach. Moreover, regularization methods yet yield further improvements.