In this thesis, we dealt with detecting surface defects on tiles under variable conditions. The research aimed to develop a robust defect detection system, analyzing the effects of different conditions on the detection accuracy and suggesting improvements. We captured a diverse dataset of images that allowed us to analyze the influence of conditions. We have artificially damaged approximately half of all tiles. We evaluated the detection system in a series of experiments, where we also tested proposed solutions to improve the robustness of the model and minimize the impact of worse conditions. We found that as the conditions deteriorate, the results also deteriorate. We developed a new model that learns in parallel on tile images and calibration sample images. This method is only a partial improvement, as the results only improve under very poor conditions. We have also implemented a color channel normalization method and a histogram matching method to correct the test images to the optimal condition using the calibration samples. We found that we were able to almost cancel out the negative effects of varying conditions using the histogram matching method.
|
