Image Registration
- PDF / 421,170 Bytes
- 8 Pages / 504.467 x 719.855 pts Page_size
- 85 Downloads / 236 Views
Image Registration Yipeng Hu, Daniel C. Alexander, and Thomy Mertzanidou UCL Centre for Medical Image Computing, Faculty of Engineering Sciences, University College London, London, UK
Synonyms Image alignment; Image fusion
Related Concepts Correspondence Image Matching Image Warping Spatial Normalization Stereo Matching
Definition Image registration aligns corresponding features of images via spatial transformations.
Background Computer vision or image processing systems often need to align multiple images of the same © Springer Nature Switzerland AG 2020 K. Ikeuchi (ed.), Computer Vision, https://doi.org/10.1007/978-3-030-03243-2_194-1
or similar scenes. In medical imaging, for example, radiologists routinely compare images of a patient acquired at different times to monitor changes. The intensity difference between two images highlights such changes, but only if the corresponding features are in the same location. However, patients’ positions in imaging devices vary between visits, so raw images never have perfect alignment. Image registration transforms or warps one image so that the important objects and the regions of interest are in the same position as in the other image. The difference image then reveals intrinsic physical changes. Figure 1 illustrates the idea. The problem becomes more challenging when the images come from different devices (inter-modality registration) or from different subjects (intersubject registration). The same problem arises in nonmedical imaging applications. Surveillance systems, for example, often need to look for differences between images at different times, for example, to subtract the background and highlight activity in a scene viewed by a security camera. Fixed cameras can wobble in the wind and produce misaligned images that require registration before the difference image provides a meaningful result. Stitching images together to create panoramas [1] also requires image registration to align the overlapping parts of the images being stitched together, as illustrated in Fig. 2. Similarly superresolution techniques [2] align multiple images of the same scene and infer sub-pixel details.
2
Image Registration
Image Registration, Fig. 1 Intra-subject brain image registration. Top left: overlaid images of the same brain from different acquisitions, one in red and one in green; bottom left: difference image of the two unaligned images;
top right: overlaid images after registration via a rigid transformation; bottom right: difference image after alignment
The theory of image registration is outlined in the following section, and an additional section has been included to introduce the recent development of image registration methods using deep learning, which share many theoretical and practical aspects of the “classical” registration approaches. The literature contains many review papers, for example, [3–8].
Theory The process of automatic image registration involves optimizing a cost function, which expresses the similarity of the two images, with respect
Data Loading...
