China
Africa
Explore chapters and articles related to this topic
Studies on Registration and Fusion of Retinal Images
Published in Rick S. Blum, Zheng Liu, Multi-Sensor Image Fusion and Its Applications, 2018
France Laliberté, Langis Gagnon
Phase correlation is a Fourier-based technique that can be used to register shifted images. Two shifted images have the same Fourier amplitude and a phase difference proportional to the translation. The cross-power spectrum (Fourier transform of the cross-correlation function) of the two images is computed, and its phase is transformed back in the spatial domain to obtain the translation. Rotation and scale can be taken into account but only at the expense of computation time or robustness.
Methods for Estimating the Optical Flow on Fluids and Deformable River Streams: -A Critical Survey
Published in Panagiotis Tsakalides, Athanasia Panousopoulou, Grigorios Tsagkatakis, Luis Montestruque, Smart Water Grids, 2018
Konstantinos Bacharidis, Konstantia Moirogiorgou, George Livanos, Andreas Savakis, Michalis Zervakis
The main advantage of implementing phase correlation in the DFT domain instead of traditional correlation in the spatial domain is that the resulted peak is sharper and at higher SNR [75]. However, the main drawbacks of phase correlation approaches relate to their conditioning on image characteristics, i.e. the filters used are heavily dependent on the signal-to-noise ratio ([55,75]).
Virtual Restoration of Antique Books and Photographs
Published in Filippo Stanco, Sebastiano Battiato, Giovanni Gallo, Digital Imaging for Cultural Heritage Preservation, 2017
Filippo Stanco, Sebastiano Battiato, Giovanni Gallo
The second part is the estimation of the displacement between the two pieces of the photographic glass plate. We use the well-known phase-correlation technique that is exploited in various motion estimation algorithms [35, 53]. According to the properties of the Fourier transform, it is possible to estimate the displacements between two images as follows:
Deep learning based tracked X-ray for surgery guidance
Published in Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 2022
K. Bamps, Stijn De Buck, Joris Ector
Although registration methodologies are well studied (Markelj et al. 2012), these methods still heavily depend on determining correct initialisation parameters to enable registration convergence (Otake et al. 2012). Different solutions have been proposed to the problem of X-ray source pose estimation. In (Mj et al. 2010), the projectionslice theorem in combination with phase correlation is used in order to estimate the transformation parameters with respect to a 3-D CT image. Despite the good results, the proposed method is not suited for real-time intraoperative setup due to the computation time. Recently, deep learning (DL) methods have shown great results in various fields. These methods can be divided into two groups, namely end-to-end and hybrid methods. In end-to-end methods, an artificial neural network is trained to estimate the camera pose parameters in one pass. An approach based on end-to-end pose estimation was put forward by (Bui et al. 2017). They used simulated X-ray images of a 3-D CAD model to train a CNN to immediately predict the pose of a mobile X-ray arm relative to the 3-D CAD model. Despite the appealing fact that feature engineering is avoided, the obtained error was substantially larger than the error obtained for state-of-the-art methods. The other group, hybrid methods, combine deep learning methods with standard registration methods. These approaches apply deep learning to sub-tasks like segmentation or anatomical landmark extraction in order to facilitate the convergence of the standard methods. In (Esteban et al. 2019). a neural network is trained to estimate the initial rigid transformation between a preoperative 3-D computed tomography (CT) volume and 2-D X-rays. This neural network is first trained on simulated 2-D X-rays from CT images to extract 2-D landmarks. Afterwards, the trained network is applied to an unseen CT volume by generating artificial 2-D X-ray images. These predicted landmarks are back-projected to the unseen CT. Due to the inaccuracies, a refinement scheme is used to adjust the positions of the 3-D back-projected landmarks. Then, the meaningful refined landmarks are used to retrain the neural network. This retrained network is finally applied to intra-operative images for pose estimation. Although, promising results are reported, the retraining phase of their method takes around 40 min. This makes the proposed automatic framework unfit for real-time use. In (Andress et al. 2018). a system is proposed that uses one multi-modality marker to register an optical see-through head mounted display to a fluoroscopy imaging system. The marker was crafted out of Pb which strongly attenuates X-ray radiation. This may result in an obstruction of the view on relevant anatomical structures or instruments. Due to the high contrast image, the marker could be detected automatically by using ARToolKit (ARt 2004).