In this work, a deep convolutional neural network is proposed to improve the registration of microtopographic data. For this purpose, different mechanical surfaces were optically measured using a confocal laser scanning microscope. A wide range of surfaces with different materials, processing methods, and topographic properties, such as isotropy and anisotropy or stochastic and deterministic features, are included. Training and testing datasets with known homographies are generated from these measurements by cropping a fixed and moving image patch from each topography and then randomly perturbing the latter. A pseudo-siamese network architecture based on the VGG Net is then used to predict these homographies. The network is trained with a supervised learning approach where the Euclidean distance between the predicted and the ground truth gives the loss function. The 4-point homography parameterization is used to improve the loss convergence. Furthermore, different amounts of image noise are added to enhance the prediction’s robustness and prevent overfitting. The effectiveness of the proposed method is evaluated through different experiments. First, the network performance is compared to intensity-based and feature-based conventional registration algorithms regarding the resulting error, the noise-robustness, and the processing speed. In addition, images from the Microsoft Common Objects in Context (COCO) dataset are used to verify the network’s generalization capability to new image types and contents. The results show that the learning-based approach offers much higher robustness regarding image noise and a much lower processing time. In contrast, conventional algorithms have a smaller registration error without image noise.
|