In this thesis, we address the problem of border detection between reflective surfaces during the process of deflectometry. In this process a striped pattern is projected on the surface, which makes it difficult to localize border with standard methods. To address this problem, we propose a new method for border localization, which uses convolutional neural network and active contours. We demonstrate the performance of our method on the task of car door border detection from images taken from the side of a car. Proposed network has the encoder decoder architecture and contains dilated convolutional layers for better pattern recognition. We show that robust fitting on segmented masks using active contours is the best way of fitting, and it reduces mean error from 5.00 to 2.67 pixels. On test set the proposed method achieves precision of 0.91, recall of 0.68 and F-score of 0.76. The method allows processing at approximately 4.29 frames per second.
|
