This thesis addresses the problem of detecting glass damage. We propose a new method that uses polarization data captured with a polarization camera, instead of relying on pictures taken with a classic camera. During the development of this method, we created the first publicly available dataset containing polarization pictures of windshield damage. The damage was annotated by hand. We pose the detection of glass damage as a semantic segmentation problem, where each pixel is classified as either healthy or damaged. We try to solve this problem using convolutional neural networks. Furthermore, we evaluate different ways of processing polarization pictures to determine the optimal processing method. The best results are offered by a model, that uses unchanged polarization images, arranged into a four channel image, which achieves a precision of 0.923, recall of 0.861 and F-score of 0.885.
|
