A brief analysis of existing algorithms to compare image-to-image, image-to-video, video-to-video
Image Similarity Measure
p167: https://link.springer.com/content/pdf/10.1007%2F978-3-642-11216-4.pdf
There is no universal similarity measure which can be used in all applications. In selecting a suitable similarity measure we find it useful to broadly divide them into two groups: Global Measure. These measures return a single similarity value which describes the overall similarity of the two input images. The global measures may be further divided into measures which require the input images to be spatially registered and those which do not require the input images to be spatially registered. Local Measures. These measures return a similarity image or map which describes the local similarity of the two input images. By definition the local similarity measures require the input images to be spatially registered.
Intensity based similarity measures – From the medical image registration field
sum of squared differences (SSD) – Thus, the lower the SSD is, the better the registered images is.
The cross-correlation and its derived measures, such as the Pearson’s correlation coefficient or correlation ratio, – Thus, the larger the crosscorrelation is, the better the registered image is.
a normalized mutual information (NMI) – The MI is based on the Shannon entropy that is computed from the joint probability distribution of the image voxel intensity.
one based on a combination of MI and gradient information, and the other one based on NMI and gradient information.
regional mutual information, and in (Loeckx et al., 2010), using the conditional mutual information.
similarity measure used is based on the Euclidean distance between the principal eigenvectors of the diffusion tensors.
minimization of the symmetrised Kullback-Leibler divergence between the Gaussian probability density functions whose covariance matrices are given by the diffusion tensors
Feature based similarity measures – From the medical image registration field
- Segmentation or feature extraction and then similarity measure
similarity measures based on the curvature have been used in surface matching
correlation ratio is considered as the similarity measure used to register sets of fibres extracted from brain white matter images.
MI is computed using the image gradient fields.
Segmentation Algorithm:
region-based techniques are: thresholding methods (Otsu, 1979; Wellner, 1993), watershed (Beucher, 1991; Grau et al., 2004), and region growing (Adams and Bischof, 1994). Usual border-based segmentation techniques include edge detectors based on image gradient (Canny, 1986; Marr and Hildreth, 1980), corner detectors, line detectors based on the Hough transform; deformable models, like active contours, usually known as snakes, (Cootes and Taylor, 1992; Gonçalves et al., 2008; Kass et al., 1988; McInerney and Terzopoulos, 1996; Xu and Prince, 1998) and level set methods (Han et al., 2009; Wang et al., 2007; Wang and Wang, 2006).
Reviews on image segmentation techniques can be found in (Gonzalez and Woods, 2008; Ma, Zhen et al., 2010; Monteiro, 2007; Zhang and Lu, 2004; Zhang, 2001).
Frequency based methodologies
The SSD and cross-correlation based similarity measures can be efficiently evaluated in the frequency domain using the Fourier transform and its properties.
