Using artificial intelligence to detect camera lens contamination

Digital image processing is a crucial task with a large number of applications. In rough environments, the camera lens gets polluted over time. This can lead to problems in subsequent image-processing tasks. To protect the camera lens, a glass plate is often placed in front of the lens to prevent it from becoming dirty. But this layer must be cleaned over time to get clear pictures for the subsequent image processing. The task of this thesis was to detect when the layer is too dirty and needs to be cleaned or replaced. For the detection, the camera images were used to train a Convolutional Neural Network (CNN). Ideally, the detection should work in different environments. Therefore, a lot of data for training is needed. To provide this data, a simulator is used, that can simulate the behavior of a polluted image and can thus generate an arbitrary number of images. By tuning the parameters of the simulator, different data sets with different pollution patterns and degrees of pollution were created. Subsequently, the simulated dirty images were used to train distinct state-of-the-art CNN architectures. The results show, that with some architectures it is possible to classify the simulated dirty images with a mean accuracy of over 80%. Furthermore, the models, trained by simulated dirty images, were applied to real- world data. The results show, that with the choice of the right data set, it is possible to classify images with real pollution patterns, by a model which is trained by simulated pollution. The best individual models are able to classify the images with a balanced accuracy of 79%.